WorksComputer Science
Renegotiable Options Fail to Prevent Hold-Up Between Two Profit-Maximizing Frontier LLM Agents
Abstract
Do contract terms that protect investment between people also protect it between profit-maximizing language-model agents? In the hold-up problem, a seller can adapt a unit to one buyer at a cost before the price is set. Once that cost is sunk, the buyer can offer almost nothing and the seller will still accept, so a seller who foresees this does not adapt. One remedy is an option: the buyer may buy at a fixed price. In theory the option stops working if the buyer can let it lapse and offer its own price instead. Yet in a laboratory experiment by Hoppe and Schmitz, this renegotiable option raised human investment from 41.5% of pairs without a contract to 67.2%. We had Claude Opus 5.5 and GPT-6.1 Sol, each told to maximize its side's profit, play a one-shot game with the same strategic structure under their four contract regimes and a reference-price control. The agents did what game theory predicts: under the renegotiable option no seller adapted, and under the option that rules out this renegotiation every seller did. In a second experiment in which sellers no longer stated their expected revenue before deciding, removing the profit instruction left investment unchanged under no contract and both options. When we changed the numbers so that the buyer would exercise the option even without adaptation, adapting only added cost, and no seller adapted: the agents did not simply adapt whenever an option was offered or whenever its price covered the cost. Under our protocol, a contract term that raised investment between people did not protect it between agents.
Introduction
A supplier asked to adapt equipment to one customer faces a familiar problem. The adaptation costs money now, is useless to anyone else, and the price is negotiated afterwards. Once the money is spent, the customer
AI assistance: The authors used Claude Code (Anthropic) during the preparation of this work. can offer a price barely above the supplier’s remaining alternative, which is nothing. Suppose adapting costs $8k and raises the unit’s value to the customer from $10k to $22k. Once the $8k is spent, a customer who offers $1k leaves the supplier at $1k − $8k = −$7k, and the supplier accepts, since rejecting leaves it at −$8k. A supplier who foresees this does not adapt, and the value the adaptation would create is lost. This is the hold-up problem, and the adaptation is a relationship-specific investment [20, 26, 10, 12].
Contracts are the standard remedy. Suppose the customer holds an option to buy at $15k. It exercises only if the unit is worth more than $15k to it, which is true after adaptation ($22k) and not before ($10k), so adapting secures the sale and the supplier nets $15k − $8k = $7k. If the customer can instead let the option lapse and offer its own price, the protection disappears: once the $8k is spent, it offers $1k, and the supplier accepts as before. Options and breach remedies (damages owed by a party that breaks the contract) protect an investment that raises the investor’s own value even when the parties can renegotiate [22, 6]. Here the investment raises the customer’s value, the case in which a contract open to renegotiation may do no better than none [4].
Laboratory subjects did not follow this prediction. In the experiment of Hoppe & Schmitz [15], the seller invested in 67.2% of pairs with a renegotiable option, against 41.5% without a contract. The authors argue that standard models of social preferences, in which people care about the other party’s payoff as well as their own, predict the same investment under the renegotiable option as without a contract, because in both the buyer is free to set the price after the investment. These models therefore do not explain the increase. They attribute it to the contract acting as a reference point, a benchmark for what each party feels entitled to [13, 9].
Firms may delegate procurement and sales to language-model agents told to maximize profit, and these agents would then trade with each other. Whoever designs the contracts they sign needs to know whether terms that work through human behavior still protect investment between such agents. In one game from our study, a seller agent (Claude Opus 5.5, Opus below) and a buyer agent (GPT-6.1 Sol, Sol below) were offered the renegotiable option with the original numbers and an option price of $15k. The agents’ prompts state amounts in thousands of dollars, so the seller’s 8 below is $8k. Asked whether to adapt, the seller answered: “Whether or not I adapt, the buyer gains more by declining the option and offering about 0, which I would accept. Adapting would cost 8 with no revenue gain, so I don’t adapt.” The buyer declined the option on the unadapted unit and offered $1k, and the seller accepted. The reply reasons as the theory says. Across all games, the renegotiable option, which raised investment in the human experiment, left investment at zero.
We compare five regimes (Table 1): four of Hoppe & Schmitz [15] and a reference-price control. The regimes are no contract (NC), a reference price (REF), a fixed price (FP), an option (OC) and a renegotiable option (OCR). OC and OCR differ only in whether the buyer can walk away from the option and offer its own price. Each game is one-shot. Its subgame-perfect prediction, found by backward induction (solving the game from the last move back, with each party maximizing its own payoff at every step), is investment under OC only (§ 2 and Figure 1).
Table 1: The five regimes. Agents invest exactly where theory predicts. The human subjects of Hoppe & Schmitz (2011) also invested in regimes where it predicts none. High-investment rate: the share of pairs in which both parties took part and the seller adapted. Human rates come from Hoppe & Schmitz (2011), from Figure 2 of their author manuscript [14], which gives the count and the number of pairs in each treatment. Their denominator is all pairs, including pairs in which a party declined to take part or to sign. Their games were one-shot laboratory games with the values , , and of our original numbers, in euros, and the same strategic structure. Each pair played once. In their NC+ treatment (no contract), as in our NC, the buyer makes a take-it-or-leave-it price offer. Agent rates are identical in all three pairings of seller and buyer models, each with 96 games per regime.
| Regime | After the seller’s adaptation decision | Prediction | Humans | Agents | |
| No contract | NC | the buyer offers a price; the seller accepts or rejects | no investment | 41.5% (22/53) | 0% |
| Reference price | REF | as NC, with $15k named as a non-binding reference price | no investment | — | 0% | |||
| Fixed price | FP | the sale takes place at $15k automatically | no investment | 23.4% (15/64) | 0% | |||
| Option | OC | the buyer may buy at $15k; otherwise no sale | investment | 90.9% (50/55) | 100% | |||
| Renegotiable option | OCR | the buyer may buy at $15k, or instead offer its own price; the seller accepts or rejects | no investment | 67.2% (43/64) | 0% |
Table 1.

Figure 1. Move 1. Entry or signing (both parties, at the same time); either declines → (0, 0), FP⋆: the buyer declines; both take part. Move 2. Seller: adapt? adapt, cost , equilibrium under OC; do not adapt, equilibrium under NC, REF, FP, OCR. Move 3, adapted: worth to the buyer. NC, REF: buyer offers 0 → seller accepts ; FP: sale at 15 ; OC: buyer exercises ; OCR: buyer lets the option lapse → offers 0 → seller accepts . Move 3, not adapted: worth to the buyer. NC, REF: buyer offers 0 → seller accepts ; FP: sale at 15 ; OC: buyer does not exercise; no sale ; OCR: buyer lets the option lapse → offers 0 → seller accepts .
Each regime was played by three seller–buyer pairings: an Opus seller with a Sol buyer, a Sol seller with an Opus buyer, and Sol with Sol. A separate module varies bargaining power, the ability to name a take-it-or-leave-it price and so keep the surplus: the buyer invests, a programmed seller is the counterpart, and either the program or the buyer names the price.
Our contributions:
Which contract terms produce investment. Only the option without renegotiation did: sellers adapted in every OC game and in no NC, REF or OCR game (§ 4.1). To check that the instruction to maximize profit does not by itself produce this play, we removed it in an experiment in which sellers no longer stated their expected revenue before deciding; sellers still adapted in every OC game and in no NC or OCR game (§ 4.3).
What a buyer would pay for an adapted unit. No seller adapted under NC or OCR, so the games never show a buyer’s response to an adapted unit, and a probe asked buyers alone. Buyers told that the seller had adapted offered at most $1k for a unit worth $22k, or $66k with tripled numbers. Had sellers accepted these offers, the extra revenue from adapting would have covered almost none of its cost (§ 4.2).
Following the payoffs rather than a one-condition rule. With the original numbers, equilibrium play agrees with the rules “adapt if ” and “adapt if ”, so a flip study drew new numbers for the option and fixed-price contracts for which the equilibrium action reverses. With a low option price the buyer buys even the unadapted unit, so adapting only adds cost; every seller reached that choice, and none adapted. The other case, in which adapting wins the sale but loses money, was reached only by Opus sellers, since all Sol sellers and some buyers declined a contract that pays nothing either way (§ 4.4).
Investment when the investor sets the price. Theory says hold-up disappears when the investor sets the price, which the seller never does in the contract games. In a game in which the buyer invests against a pricing program, Opus and Sol invested only when they set the price, at a payoff-maximizing level in all 768 of their decisions (§ 4.5).
Game and predictions
The hold-up game. A seller can adapt a unit for one buyer at a cost . The unit is worth to the buyer without adaptation and with it, and nothing to anyone else. The buyer observes whether the seller adapted, but no court can verify it, so a price cannot be made contingent on adaptation. With the values of Hoppe & Schmitz [15], , , (euros there, thousand US dollars here), and adapting creates a net surplus of 4. Every game starts with both parties deciding at the same time whether to take part or to sign the contract on offer. If either declines, both earn 0. Then the seller decides whether to adapt, which is its investment, and the regime determines what follows (Figure 1).
Regimes. We use four regimes of Hoppe & Schmitz [15] and one control, REF, with their contract price .
Figure 1: The game with the original numbers (, , , ) and its equilibrium in each regime. The moves are made in order, from top to bottom. Payoffs are (seller, buyer) in thousand dollars. Each box gives the equilibrium choices of the last move in each regime; ⋆ marks the outcome reached in equilibrium. Under FP the seller would not adapt, leaving the buyer , so the buyer declines to sign. If the seller would reject an offer of 0, the buyer offers 1 instead, which changes no investment or signing prediction. The entry stage also has equilibria in which both parties decline; the figure shows the one in which every party takes part unless that lowers its payoff.
NC (no contract). After the adaptation decision, the buyer offers any whole-number price up to the unit’s value, and the seller accepts or rejects. This is their NC+ treatment (no contract, with the participation decision of their FP, OC and OCR treatments), the source of their 41.5%.
REF. NC plus one sentence stating that $15k is a commonly quoted, non-binding reference price. It tests whether naming the price without binding anyone changes play, since contracts can act on people as reference points [9].
FP (fixed price). The sale takes place automatically at after the adaptation decision.
OC (option contract). The buyer may buy at after observing the adaptation decision. If it does not, no sale can ever take place.
OCR (option with renegotiation). As OC, but if the buyer does not exercise the option, it may offer a new price, which the seller accepts or rejects.
Predictions. Backward induction (Figure 1) gives the predictions in Table 1. Without a contract, or with only the reference price, the buyer offers 0 whatever the seller did, so adapting only costs the seller 8. Under FP the price is fixed, so the seller never adapts. The buyer would then get and refuses to sign. Under OC the buyer exercises when the unit is worth more than , which with happens only after adaptation, so the seller adapts and earns . Under OCR the buyer lets the option lapse and offers 0 whatever the seller did, so the seller does not adapt. A game matches the prediction when every entry or signing, adaptation and exercise decision taken is an equilibrium choice, one made in some subgame-perfect equilibrium in which each party takes part unless that lowers its payoff (Figure 1), at the point of the game where it is made. Both 0 and 1 are equilibrium offers: a seller offered 0 earns 0 whether it accepts or rejects, so accepting is a best reply, and a buyer facing a seller who would reject 0 offers 1, the smallest positive price. Some entry and signing decisions are ties of the same kind, at which both choices pay the deciding party the same given the other’s equilibrium play, and either choice there counts as equilibrium play. Prices and their acceptance are not part of the criterion; we report them separately.
Experimental design
Studies. Table 2 lists the five studies. The contract games are the main study; each of the other four answers a question the contract games leave open.
| Study | Question | Design | Section |
| Contract games | Which of the five regimes produce investment? | 96 scenarios × 5 regimes × 3 pairings = 1440 games | 4.1 |
| Buyer probe | What would a buyer offer for an adapted unit? If no seller adapts under NC or OCR, the games never show it. | 96 scenarios × 2 regimes × 2 adaptation states × 2 buyer models = 768 probe contexts | 4.2 |
| Without the profit instruction | Does the instruction to maximize profit by itself produce textbook play? | 96 scenarios × 3 regimes × 2 pairings × 2 prompts (with and without the instruction) = 1152 games; 96 × 2 regimes × 2 buyer models × 2 prompts = 768 probe contexts | 4.3 |
| Flip study | Do agents play the equilibrium or a one-condition rule (adapt if ; adapt if )? | 32 scenarios with new numbers × 2 contracts × 2 prices × 3 pairings = 384 games | 4.4 |
| Bargaining power | Does the investor invest when it sets the price, as theory predicts? | 96 sets of values by investment level × 2 price setters × 2 levels of rule detail × 2 models = 768 decisions by Opus and Sol; the same for two smaller models | 4.5 |
Table 2. The five studies. Design gives the factors whose product is the size.
Agents. Claude Opus 5.5 (Anthropic API, provider-default adaptive reasoning) and GPT-6.1 Sol (Azure OpenAI, reasoning effort medium) played both roles (other settings in Appendix § 10). Each agent’s system prompt named its principal, instructed it to maximize that principal’s own monetary profit, and gave the full rules (Appendix § 9). Every game decision was a JSON reply with a reason; the prompt asked for at most 40 words. A reply that was not valid JSON with the required fields and values received one repair request; a call that timed out or was refused by the provider was retried and then resent unchanged, except that in the experiment without the profit instruction a refused reply received the repair request (Appendix § 10). At entry or signing the buyer also stated its probability that the seller would adapt. At the adaptation decision the seller stated its expected revenue with and without adapting. Each agent kept its own conversation and saw only the other party’s moves (entry or signing, adaptation, exercise, price), never its reasons or its model. Each game was one-shot. After each contract-game and flip-study game, both parties answered a short comprehension question. Claude Haiku 5.5 and GPT-5.6 Luna appear only in the bargaining-power module (§ 4.5).
Scenarios. The 96 scenarios come from 32 business domains with three sites each (for example, a seawater treatment unit adapted to filter collars embedded in an aquarium’s tank walls; Table 6). Each scenario uses one of two wordings of the prompts and one of two parameter sets: the original numbers, or , , and tripled. Both sets give the same investment and signing predictions. Along the scenarios ordered by domain and site, wordings alternate one by one and parameter sets two by two, so each of the four combinations has 24 scenarios and every domain has both wordings and both parameter sets (Appendix § 10). Wording B rephrases each prompt of wording A (for example, “customize” for “adapt”; Appendix § 9). Every scenario was played in all five regimes and three pairings, giving 1440 games. We write a pairing as seller>buyer, so Opus>Sol has an Opus seller and a Sol buyer. The pairings place each model in each role against Sol, the less expensive model.
Table 6: Scenario domains. Each domain has three scenarios, one per site (North, West, East). The buyer is the domain’s institution at that site (for example, North Aquarium), it buys the service, and the adapted part is embedded in the named structure.
| Dom. | Domain | Service | Adapted part | Embedded in |
| 1 | Aquarium | seawater treatment | filter interface collars | tank walls |
| 2 | Athletics Center | performance timing | timing-system anchors | running lanes |
| 3 | Botanical Garden | mist irrigation | spray-head connectors | greenhouse frames |
| 4 | Cable Car | cable inspection | inspection-head sockets | tower crossbeams |
| 5 | Campground | utility servicing | utility connection plates | camping pads |
| 6 | Concert Hall | stage-lighting service | light fixture sockets | ceiling trusses |
| 7 | Convention Center | visitor assistance | information kiosk docks | foyer columns |
| 8 | Cruise Terminal | boarding service | gangway alignment sockets | quay walls |
| 9 | Equestrian Center | horse rehabilitation | gait sensor plates | exercise tracks |
| 10 | Ferry Port | vessel charging | charging interface plates | berth foundations |
| 11 | Film Studio | motion capture | tracking beacon docks | soundstage walls |
| 12 | Funeral Park | grounds irrigation | irrigation-control couplers | buried conduits |
| 13 | Golf Course | turf monitoring | soil-probe sleeves | fairways |
| 14 | Ice Rink | ice condition monitoring | temperature probe sleeves | rink floors |
| 15 | Marina | berth servicing | mooring sockets | floating pontoons |
| 16 | Museum | exhibit conservation | climate-control couplings | display cases |
| 17 | Observatory | telescope maintenance | alignment reference sockets | telescope piers |
| 18 | Planetarium | dome illumination | light-engine mounts | dome supports |
| 19 | Quarry | rock removal | cable anchor sockets | rock faces |
| 20 | Racecourse | track timing | timing detector anchors | track edges |
| 21 | Radio Station | transmitter maintenance | test interface plates | antenna bases |
| 22 | Rail Station | passenger guidance | guidance beacon sockets | platform walls |
| 23 | Reservoir | water quality monitoring | sampling-head collars | intake structures |
| 24 | Ski Resort | snowmaking | water-pressure connectors | mountain trails |
| 25 | Solar Farm | panel inspection | inspection rail sockets | panel supports |
| 26 | Stadium | sports broadcasting | camera anchor plates | grandstand walls |
| 27 | Theme Park | ride inspection | vibration-monitoring ports | ride foundations |
| 28 | Toll Road | road condition monitoring | pavement sensor sleeves | road surfaces |
| 29 | University | lecture interpretation | interpreter console docks | lecture benches |
| 30 | Water Park | pool filtration | filter cartridge connectors | pool plumbing |
| 31 | Wind Farm | turbine inspection | inspection tether anchors | turbine foundations |
| 32 | Zoo | animal habitat monitoring | sensor docking plates | enclosure walls |
Table 6. Scenario domains. Each domain has three scenarios, one per site (North, West, East). The buyer is the domain’s institution at that site (for example, North Aquarium), it buys the service, and the adapted part is embedded in the named structure.
Buyer probe. If sellers do not adapt under NC or OCR, the games never show how a buyer responds to an adapted unit, which is what a seller must anticipate. The probe asks buyers directly. A separate buyer agent received the game’s buyer system prompt with no earlier turns, followed by the requests a game buyer receives after the adaptation decision. These state that both parties had entered (NC) or signed (OCR) and that the seller had or had not adapted. Under NC it named a price; under OCR it decided whether to exercise the option and, if the option lapsed, named a price. This gives 768 contexts (96 scenarios, two regimes, both adaptation states, Opus and Sol buyers).
Outcomes. Amounts are in thousands of US dollars on the original scale: amounts from scenarios with tripled numbers are divided by three, so an offer of 1 there is reported as 0.33. The main outcome is high investment: both parties took part and the seller adapted. For FP, where the equilibrium seller never adapts, we also report whether the buyer signs. A decision with no valid reply after the repair ends the game and counts as no investment. Most outcomes are identical in every game of a condition, so we report them as counts, with the planned tests alongside.
Inference. Each planned contrast compares two conditions within the same scenario, and the domain is the resampled unit. With one play per cell, each -value tests the contrast against sign-symmetric variation across the 32 domains, so it describes these business framings and their wordings (in the flip study and the bargaining-power module, also their parameters), not other models or prompts.
Each -value comes from a sign-flip test over the 32 domain totals with 9,999 random sign assignments, so the smallest attainable is , reached in every contrast here in which the effect has the same sign in every domain. Holm correction is applied within each family of planned contrasts, defined in § 4.1, § 4.3, § 4.4 and § 4.5 (test details in Appendix § 11). Results for the two smaller models are secondary and unadjusted. For all five studies (the contract games, the buyer probe, the experiment without the profit instruction, the flip study and the bargaining-power module), hypotheses, prompts, code and the planned analyses were fixed before their data were collected. Analyses marked additional were not among the planned analyses. The ancillary files contain the prompts, the record of every game, and the outcome and text of every invalid or repaired reply.
Results
Investment and signing match the equilibrium in every regime
Under our protocol, the two agents did not reproduce the human increase under the renegotiable option (Table 1). OCR raised high investment from 41.5% to 67.2% of pairs in the experiment of Hoppe & Schmitz [15]; between the two agents it changed nothing.
Investment. Only the non-renegotiable option produced investment, and it did so in every game. The seller adapted in all 288 OC games and in none of the 864 NC, REF and OCR games. The four planned contrasts (OC–NC, OC–OCR, OCR–NC, OCR–REF) average the Opus>Sol and Sol>Opus pairings within a scenario, so that each model enters once in each role. OC minus NC and OC minus OCR are +1.00 in every scenario (Holm-adjusted , the smallest attainable for this family of four; Appendix § 11), and OCR minus NC and OCR minus REF are 0 in every scenario. Sol>Sol gives the same values. Every adaptation and exercise decision was the equilibrium choice, and at each of them the equilibrium choice paid strictly more than the alternative. The outcomes are identical for both wordings and both parameter sets.
Entry and signing. Both parties entered or signed in all 1152 NC, REF, OC and OCR games, and every one ended in a sale; under FP, every seller signed and every buyer refused. Every entry and signing decision was an equilibrium choice. At 1728 of the 2880 decisions, the choice made paid the deciding party strictly more than the alternative. The other 1152 were seller decisions that can be ties. The 288 under FP are ties in every equilibrium: the buyer refuses, so the seller earns 0 whether or not it signs. The 864 under NC, REF and OCR are ties only in the equilibrium in which the buyer offers 0, since a seller who will not adapt then earns 0 either way; if the buyer offers 1, entering pays strictly more. The tie count of 1152 is therefore an upper bound (decision counts in Appendix § 10).
Prices. In NC, REF and OCR games the buyer offered almost nothing. Every offer was 0 or 1 on the scenario’s own scale, the two equilibrium offers, except 2 offers on tripled scenarios, of which the larger was 5 (1.67 on the original scale).
Beliefs and comprehension. At entry, buyers put the probability that the seller would adapt at 0.99 on average under OC and at most 0.05 on average in each other regime. All comprehension answers were correct.
Under NC, REF and OCR the equilibrium coincides with the textbook hold-up answer, no investment, and under OC with the textbook answer for an option contract, investment. These games therefore cannot tell whether the agents follow the stated payoffs or recall that answer; the flip study (§ 4.4) separates equilibrium play from two one-condition rules for the option where play reached the reversed decisions.
Buyers would take the value of an investment
Every buyer in the probe offered at most $1k, even for an adapted unit worth 66k with tripled numbers). Each offer was 0 or 1 on the scenario’s own scale, and under OCR no buyer exercised the option. Call the offer for the adapted unit minus the offer for the same unit unadapted the extra revenue from adapting. Had sellers accepted these offers, the cost of adapting would have exceeded this extra revenue by 7.8–8.0 on average for each buyer model and regime. That is nearly the full cost of 8.
Sellers in the games. Sellers in the NC, REF and OCR games expected as much. Their stated expected gain from adapting, the expected revenue if adapted minus that if not (an additional summary), averaged 0.00–0.29 in each of these regimes, against 15 under OC.
Removing the profit instruction changed no investment choice
Without the instruction to maximize profit, every seller still adapted under OC and none adapted under NC or OCR. The instruction to maximize profit might by itself produce textbook play. A separate experiment therefore compared the system prompt above (arm I) with the same prompt minus its goal sentence (arm N; Appendix § 9). In both arms the adaptation request asked only for a reason and the decision, without the expected revenues, so that no request in arm N asked the seller to work out what adapting would earn. Each arm played the NC, OC and OCR games on all 96 scenarios in the Opus>Sol and Sol>Opus pairings (1152 games in total), and repeated the probe with the adapted unit under NC and OCR for both buyer models (768 contexts). Its planned family has five contrasts in high investment: OC minus OCR in arm N, N minus I in each of the three regimes, and OCR minus NC in arm N.
In both arms high investment was 0% under NC and OCR and 100% under OC, and the difference between the arms was zero in every scenario in each regime. In arm N, OC minus NC was +1.00 in every scenario (Holm-adjusted , the smallest attainable for this family of five). All secondary contrasts were 0 in every scenario: (N−I in OCR)−(N−I in NC), (N−I in OC)−(N−I in OCR), N−I in each regime and pairing, and probe N−I in OCR exercise and in offers above 1. In the probe, no buyer in either arm exercised the option under OCR, and every offer was 0 or 1 on the scenario’s own scale. After every game both parties answered the comprehension question, all correctly, and every decision ended with a valid reply, 1 of them after a repair request. Without the instruction, Sol sellers declined to take part in 6 of their 96 NC games. Each decline is a tie, and so equilibrium play, in the equilibrium in which the buyer offers 0, where a seller who will not adapt earns 0 either way; the Opus buyers offered 1 in all 90 games of that pairing in which the seller entered, and against an offer of 1 entering pays strictly more. With the instruction, Opus sellers declined to sign 1 of their 96 OCR games, also a tie only in the equilibrium in which the buyer offers 0. Across both arms, 2 offers in NC and OCR games exceeded 1 on the scenario’s own scale, both on tripled scenarios. In arm N, wording B’s rules still state the payoffs as profit, and the comprehension question after each game asks for a party’s profit under both wordings; in each arm, high investment was the same under both wordings, 100% under OC and 0% under NC and OCR (96 games per wording, regime and arm).
Every reversed decision that play reached followed the new numbers
With the original numbers, the equilibrium under the option agrees with two simple rules. The cost rule adapts if the option price exceeds the cost (). The value rule adapts if the option price exceeds the unit’s unadapted value (), the condition under which the buyer does not exercise on the unadapted unit. With , and , both rules and the equilibrium say adapt. We take these rules as the textbook answer: a seller that recalls that an option protects investment, or checks only one of the two conditions, plays by one of them. Except where equals , or , the equilibrium seller adapts only when both hold, , so a price with separates it from the cost rule, and a price with separates it from the value rule (Figure 2).

Figure 2. The option price that separates each simple rule from the equilibrium. Each line orders the cost , the unadapted value , the option price and the adapted value (thousand dollars, not to scale) for the original game and for two blocks of Table 7. In kind A the buyer exercises even on the unadapted unit, so adapting only adds cost; in kind B adapting secures the sale at a loss.
The flip study keeps the OC and FP rules, both wordings and the three pairings. Each of 32 scenarios, one per domain, receives its own integer parameters, drawn independently of the original numbers (Table 7); we call a scenario with its parameters a block. Each block is played in four conditions:
| Block | Scenario | Wording | Kind | OC high | OC flipped | FP high | FP low | |||
| 1 | E036 | A | A | 8 | 14 | 42 | 16 | 10 | 19 | 7 |
| 2 | E013 | B | A | 9 | 24 | 58 | 47 | 17 | 39 | 12 |
| 3 | E017 | A | B | 30 | 17 | 63 | 57 | 21 | 32 | 8 |
| 4 | E083 | B | B | 18 | 13 | 55 | 37 | 16 | 23 | 2 |
| 5 | E037 | A | A | 5 | 13 | 50 | 40 | 10 | 18 | 10 |
| 6 | E034 | B | A | 13 | 20 | 37 | 28 | 15 | 24 | 15 |
| 7 | E095 | A | B | 18 | 14 | 53 | 32 | 16 | 18 | 9 |
| 8 | E048 | B | B | 30 | 15 | 51 | 49 | 28 | 21 | 2 |
| 9 | E058 | A | A | 6 | 11 | 26 | 18 | 9 | 14 | 2 |
| 10 | E030 | B | A | 5 | 12 | 59 | 34 | 7 | 25 | 9 |
| 11 | E012 | A | B | 21 | 11 | 53 | 51 | 15 | 21 | 3 |
| 12 | E076 | B | B | 19 | 13 | 43 | 22 | 17 | 19 | 11 |
| 13 | E004 | A | A | 8 | 17 | 61 | 43 | 10 | 23 | 8 |
| 14 | E053 | B | A | 10 | 18 | 59 | 53 | 14 | 33 | 5 |
| 15 | E069 | A | B | 15 | 8 | 27 | 19 | 12 | 10 | 5 |
| 16 | E082 | B | B | 28 | 16 | 64 | 35 | 22 | 38 | 9 |
| 17 | E073 | A | A | 6 | 10 | 26 | 12 | 8 | 13 | 2 |
| 18 | E046 | B | A | 8 | 18 | 36 | 26 | 12 | 23 | 15 |
| 19 | E050 | A | B | 27 | 17 | 50 | 32 | 19 | 22 | 7 |
| 20 | E065 | B | B | 18 | 9 | 47 | 28 | 12 | 17 | 7 |
| 21 | E002 | A | A | 7 | 12 | 24 | 19 | 10 | 14 | 2 |
| 22 | E088 | B | A | 10 | 14 | 32 | 19 | 12 | 21 | 10 |
| 23 | E060 | A | B | 30 | 12 | 56 | 39 | 19 | 30 | 3 |
| 24 | E094 | B | B | 27 | 18 | 63 | 34 | 23 | 36 | 7 |
| 25 | E009 | A | A | 14 | 25 | 45 | 32 | 23 | 31 | 19 |
| 26 | E007 | B | A | 10 | 22 | 48 | 39 | 13 | 33 | 20 |
| 27 | E064 | A | B | 31 | 18 | 62 | 39 | 24 | 20 | 8 |
| 28 | E003 | B | B | 25 | 10 | 58 | 53 | 13 | 14 | 6 |
| 29 | E047 | A | A | 9 | 16 | 64 | 51 | 12 | 29 | 12 |
| 30 | E016 | B | A | 7 | 11 | 44 | 16 | 9 | 24 | 4 |
| 31 | E028 | A | B | 16 | 9 | 44 | 40 | 12 | 18 | 3 |
| 32 | E066 | B | B | 25 | 13 | 46 | 36 | 15 | 18 | 4 |
Table 7. Parameter blocks of the flip study (thousand US dollars): cost , unadapted value , adapted value , and the contract price in each condition. Block uses the listed scenario from domain (Table 6).
OC, high price: , so the seller adapts, as in the original.
OC, flipped price: the equilibrium is not to adapt. Half the blocks are of kind A, with , where the cost rule fails; the other half are of kind B, with , where the value rule fails.
FP, high price: , so the buyer refuses, as in the original. The price is also below , so a buyer that expected adaptation half the time would sign.
FP, low price: , so the buyer signs and the seller does not adapt.
Every inequality holds by at least 2 thousand dollars, and each simple rule gets both OC conditions right in only 16 of the 32 blocks. OCR has no reversed condition, since its equilibrium is no investment at any contract price.
All 384 games, 192 in flipped conditions and 192 controls, were equilibrium play (Table 3). Sellers adapted under every high option price. Sellers signed every fixed-price contract, buyers refused every high fixed price and signed every low one, and no seller then adapted. All six planned contrasts (in each pairing, OC high minus OC flipped on high investment and FP low minus FP high on buyer signing) are +1.00 (Holm-adjusted , the smallest attainable ). Every adaptation and exercise decision was the equilibrium choice, and at each of them the equilibrium choice paid strictly more than the alternative. At 576 of the 768 entry and signing decisions, the choice made paid strictly more; at the other 192, the seller’s signing under the high fixed price and both parties’ signing at a kind-B price, both choices paid the same (Appendix § 10). All comprehension answers were correct, 3 of them after a timed-out request was resent unchanged.
| Condition | Equilibrium | Original action | Reached | Outcome | Equilibrium play |
| OC, high price | adapt | adapt | 32, 32, 32 | adapted 100% | 100% |
| OC, flipped price | do not adapt | adapt | 25, 16, 16 | adapted 0% | 100% |
| FP, high price | buyer refuses | buyer refuses | 0, 0, 0 | buyer signed 0% | 100% |
| FP, low price | buyer signs | buyer refuses | 32, 32, 32 | buyer signed 100% | 100% |
Table 3. Wherever play reached a reversed decision, agents followed the new equilibrium. Reached: games that reached the adaptation decision, of 32 per pairing, in Opus>Sol, Sol>Opus and Sol>Sol. In OC with the flipped price these are the 16 kind-A games in every pairing and, with an Opus seller only, 9 kind-B games (Table 4). Outcome and equilibrium play: shares over all 32 games per condition and pairing, identical in all three pairings. Adapted: both parties signed and the seller adapted. Original action: the equilibrium action under the original numbers. Equilibrium play: every entry, signing, adaptation and exercise decision taken was an equilibrium choice at the point where it was made (§ 2).
No seller adapted at a reversed adaptation decision. The kind-A games are the main test: both parties signed every one, and none of the 48 sellers, 16 per pairing, adapted, which rules out the cost rule for both seller models. The planned contrasts compare high investment over all games of a condition, including games that ended at signing, so this count over the games that reached the adaptation decision is an additional analysis.
| Pairing | Seller signed | Buyer signed | Both signed | Seller adapted |
| Opus>Sol | 15 | 10 | 9 | 0 |
| Sol>Opus | 0 | 16 | 0 | 0 |
| Sol>Sol | 0 | 9 | 0 | 0 |
Table 4. Kind-B games with the flipped option price, by pairing (16 games each; an additional analysis). Both signed: the games that reached the adaptation decision.
The value rule is tested only for Opus sellers (Table 4). In a kind-B block the seller does not adapt and the buyer does not buy, so both parties earn 0 whether or not they sign, and either signing choice is equilibrium play. No Sol seller signed a kind-B contract: in all 128 of their option-contract games, Sol sellers signed exactly when the price exceeded the cost (an additional analysis), a pattern that also fits the rule “sign only if the price exceeds the cost”. The 9 kind-B games in which both parties signed, all with an Opus seller, are therefore the only test of the value rule, and in none did the seller adapt. For the same reason, the planned OC contrast in kind-B blocks mostly reflects signing rather than adaptation. For Sol sellers the value rule is therefore untested.
What the fixed-price flip does not separate. Under the fixed price, buyers signed exactly when the price was below the unit’s unadapted value to them. A worst-case signing rule gives the same choices: a buyer that assumes the seller will not adapt, whatever the seller’s incentive, signs only if the price is below the unadapted value. Under a fixed price the equilibrium seller never adapts, so the two rules coincide, and the fixed-price flip is consistent with both.
Opus and Sol invest as buyers only when they set the price
Theory says hold-up disappears when the investor sets the price, because it then keeps the value its investment creates. In the contract games the seller invests and never names a price; the buyer names it under NC, REF and the renegotiation of OCR, and the contract fixes it otherwise, so the investor never holds this power. This module makes the buyer agent the investor and lets it set the price. The agent chooses an investment level from 0 to 3 for one of 96 value schedules (Table 8); its counterpart, the seller, is an automated pricing program that maximizes the seller’s revenue. Because the program’s reaction follows from the stated rules, the agent’s best level is known exactly in every schedule, and the module tests only the agent’s own foresight, not its forecast of another agent. With a deal the agent earns its level’s value minus the price and the level’s cost; without one, minus the cost.
The two protocols are named for the side that makes the take-it-or-leave-it offer. When the program sets the price, it names a price after seeing the level, and the agent accepts or rejects. Its best price, the value minus 0.5, leaves the agent half a thousand dollars, so the agent should invest nothing. When the agent sets the price, as the buyer does in the contract games but now as the investor, the program accepts any price of at least 0.5, so the agent should choose a payoff-maximizing level (price grid, tie rule and the scoring of invalid replies in Appendix B). The agent saw one of two information conditions: the rules only, or the rules with an explicit statement of the program’s reaction. We write for the lowest payoff-maximizing level of a schedule (Table 8). With its level, the agent also stated, for every level, the price it expected the program to name or planned to name itself. Regret, the payoff given up relative to the best choice, is the best attainable payoff minus the realized net payoff. Two smaller models, Claude Haiku 5.5 and GPT-5.6 Luna, also play the module, which shows how much an agent loses when it misjudges the program’s reaction (Table 5).
| Model | Information | Regret | Investment loss | Price loss | Failure loss |
| GPT-5.6 Luna | rules only | 88.9 | 19.4 | 69.5 | 0.0 |
| GPT-5.6 Luna | with reaction | 37.2 | 3.4 | 33.8 | 0.0 |
| Claude Haiku 5.5 | rules only | 31.8 | 10.1 | 18.4 | 3.3 |
| Claude Haiku 5.5 | with reaction | 5.1 | 1.8 | 0.0 | 3.3 |
Table 5.
| ID | Dom. | Site | W. | Shape | ID | Dom. | Site | W. | Shape | ||||||||||||
| E001 | 11 | E | A | dim. | 80 | 145 | 167.5 | 172.5 | 15 | 2 | E049 | 8 | N | A | thr. | 120 | 130 | 175 | 185 | 20 | 2 |
| E002 | 21 | W | B | dim. | 100 | 137.5 | 167.5 | 177.5 | 20 | 2 | E050 | 19 | E | A | thr. | 110 | 115 | 190 | 210 | 25 | 2 |
| E003 | 28 | E | A | dim. | 90 | 127.5 | 157.5 | 167.5 | 15 | 2 | E051 | 4 | W | B | thr. | 100 | 107.5 | 160 | 167.5 | 20 | 2 |
| E004 | 13 | N | A | dim. | 110 | 130 | 145 | 150 | 10 | 2 | E052 | 20 | W | B | thr. | 90 | 95 | 175 | 195 | 20 | 2 |
| E005 | 10 | W | B | dim. | 100 | 195 | 242.5 | 257.5 | 25 | 2 | E053 | 14 | N | A | thr. | 100 | 102.5 | 142.5 | 152.5 | 10 | 2 |
| E006 | 14 | E | A | dim. | 110 | 155 | 177.5 | 185 | 20 | 2 | E054 | 7 | W | A | thr. | 120 | 132.5 | 220 | 232.5 | 20 | 2 |
| E007 | 26 | E | A | dim. | 80 | 145 | 167.5 | 172.5 | 25 | 1 | E055 | 19 | W | B | thr. | 80 | 85 | 130 | 135 | 10 | 2 |
| E008 | 9 | W | B | dim. | 80 | 132.5 | 152.5 | 157.5 | 15 | 2 | E056 | 2 | W | B | thr. | 120 | 125 | 150 | 155 | 10 | 2 |
| E009 | 25 | E | B | dim. | 80 | 127.5 | 165 | 180 | 15 | 2 | E057 | 30 | W | A | thr. | 120 | 125 | 225 | 250 | 25 | 2 |
| E010 | 23 | E | A | dim. | 110 | 135 | 155 | 162.5 | 10 | 2 | E058 | 9 | N | A | thr. | 90 | 97.5 | 127.5 | 135 | 10 | 2 |
| E011 | 25 | W | A | dim. | 120 | 157.5 | 177.5 | 182.5 | 10 | 2 | E059 | 1 | W | B | thr. | 80 | 85 | 120 | 125 | 10 | 2 |
| E012 | 11 | W | B | dim. | 100 | 165 | 187.5 | 192.5 | 25 | 1 | E060 | 23 | N | A | thr. | 90 | 92.5 | 132.5 | 142.5 | 10 | 2 |
| E013 | 2 | E | A | dim. | 110 | 135 | 145 | 147.5 | 10 | 1 | E061 | 29 | N | B | thr. | 80 | 90 | 155 | 165 | 25 | 2 |
| E014 | 23 | W | B | dim. | 90 | 127.5 | 157.5 | 167.5 | 20 | 2 | E062 | 11 | N | A | thr. | 110 | 120 | 165 | 175 | 15 | 2 |
| E015 | 12 | E | A | dim. | 80 | 190 | 230 | 237.5 | 25 | 2 | E063 | 17 | N | A | thr. | 110 | 120 | 190 | 200 | 20 | 2 |
| E016 | 30 | E | B | dim. | 80 | 157.5 | 220 | 242.5 | 25 | 2 | E064 | 27 | N | B | thr. | 120 | 132.5 | 220 | 232.5 | 25 | 2 |
| E017† | 3 | W | B | dim. | 90 | 115 | 125 | 127.5 | 10 | 1 | E065 | 20 | E | B | lin. | 90 | 100 | 112.5 | 127.5 | 10 | 3 |
| E018 | 32 | N | A | dim. | 120 | 160 | 175 | 177.5 | 15 | 1 | E066 | 32 | E | B | lin. | 80 | 110 | 145 | 185 | 20 | 3 |
| E019 | 12 | W | B | dim. | 100 | 137.5 | 157.5 | 162.5 | 10 | 2 | E067 | 29 | W | B | lin. | 100 | 120 | 137.5 | 155 | 15 | 3 |
| E020 | 21 | E | B | dim. | 110 | 185 | 222.5 | 235 | 20 | 2 | E068 | 22 | E | B | lin. | 80 | 105 | 125 | 140 | 10 | 3 |
| E021 | 6 | N | B | dim. | 90 | 167.5 | 230 | 252.5 | 25 | 2 | E069 | 15 | E | B | lin. | 110 | 147.5 | 192.5 | 242.5 | 25 | 3 |
| E022 | 31 | E | B | dim. | 90 | 150 | 180 | 190 | 20 | 2 | E070 | 28 | N | A | lin. | 120 | 137.5 | 157.5 | 180 | 15 | 3 |
| E023† | 3 | N | A | dim. | 90 | 142.5 | 162.5 | 167.5 | 20 | 1 | E071 | 22 | N | A | lin. | 90 | 122.5 | 152.5 | 182.5 | 15 | 3 |
| E024 | 16 | W | A | dim. | 90 | 165 | 202.5 | 215 | 25 | 2 | E072 | 13 | E | B | lin. | 90 | 107.5 | 127.5 | 150 | 15 | 3 |
| E025 | 25 | N | B | dim. | 110 | 172.5 | 222.5 | 242.5 | 20 | 2 | E073 | 17 | W | B | lin. | 120 | 150 | 185 | 225 | 20 | 3 |
| E026 | 9 | E | B | dim. | 110 | 155 | 177.5 | 185 | 15 | 2 | E074 | 24 | E | A | lin. | 120 | 135 | 145 | 155 | 10 | 1 |
| E027 | 24 | W | A | dim. | 80 | 167.5 | 197.5 | 202.5 | 25 | 2 | E075 | 20 | N | A | lin. | 110 | 172.5 | 220 | 260 | 25 | 3 |
| E028 | 31 | W | A | dim. | 120 | 145 | 155 | 157.5 | 10 | 1 | E076 | 12 | N | B | lin. | 120 | 170 | 207.5 | 237.5 | 20 | 3 |
| E029 | 13 | W | A | dim. | 120 | 165 | 187.5 | 195 | 20 | 2 | E077† | 3 | E | A | lin. | 80 | 102.5 | 120 | 135 | 15 | 2 |
| E030 | 10 | E | B | dim. | 90 | 112.5 | 122.5 | 127.5 | 10 | 1 | E078 | 31 | N | B | lin. | 110 | 127.5 | 145 | 160 | 10 | 3 |
| E031 | 16 | N | A | dim. | 100 | 137.5 | 167.5 | 177.5 | 15 | 2 | E079 | 8 | W | B | lin. | 120 | 147.5 | 180 | 217.5 | 15 | 3 |
| E032 | 30 | N | B | dim. | 110 | 162.5 | 182.5 | 187.5 | 20 | 1 | E080 | 14 | W | B | lin. | 120 | 152.5 | 182.5 | 212.5 | 25 | 3 |
| E033 | 7 | E | B | thr. | 90 | 95 | 130 | 135 | 10 | 2 | E081 | 17 | E | A | lin. | 90 | 107.5 | 125 | 140 | 10 | 3 |
| E034 | 6 | E | A | thr. | 80 | 87.5 | 122.5 | 130 | 15 | 2 | E082 | 16 | E | B | lin. | 120 | 160 | 190 | 215 | 20 | 3 |
| E035 | 32 | W | B | thr. | 90 | 92.5 | 125 | 130 | 10 | 2 | E083 | 4 | N | A | lin. | 90 | 110 | 125 | 137.5 | 10 | 3 |
| E036 | 1 | N | A | thr. | 80 | 92.5 | 180 | 192.5 | 20 | 2 | E084 | 10 | N | A | lin. | 120 | 182.5 | 230 | 270 | 25 | 3 |
| E037 | 5 | N | B | thr. | 120 | 140 | 215 | 235 | 25 | 2 | E085 | 18 | N | A | lin. | 100 | 135 | 167.5 | 197.5 | 20 | 3 |
| E038 | 27 | E | B | thr. | 120 | 135 | 245 | 260 | 25 | 2 | E086 | 28 | W | B | lin. | 120 | 155 | 187.5 | 217.5 | 20 | 3 |
| E039 | 21 | N | B | thr. | 90 | 95 | 117.5 | 122.5 | 10 | 2 | E087 | 1 | E | A | lin. | 100 | 125 | 150 | 172.5 | 15 | 3 |
| E040 | 6 | W | B | thr. | 120 | 135 | 190 | 205 | 15 | 2 | E088 | 22 | W | A | lin. | 80 | 125 | 165 | 205 | 25 | 3 |
| E041 | 2 | N | A | thr. | 90 | 95 | 155 | 170 | 15 | 2 | E089 | 26 | W | A | lin. | 80 | 95 | 105 | 115 | 10 | 1 |
| E042 | 27 | W | A | thr. | 90 | 100 | 145 | 155 | 15 | 2 | E090 | 15 | W | A | lin. | 100 | 137.5 | 165 | 187.5 | 15 | 3 |
| E043 | 5 | E | A | thr. | 90 | 95 | 170 | 190 | 25 | 2 | E091 | 5 | W | B | lin. | 110 | 150 | 180 | 205 | 20 | 3 |
| E044 | 4 | E | B | thr. | 80 | 85 | 185 | 210 | 20 | 3 | E092 | 19 | N | A | lin. | 90 | 115 | 140 | 162.5 | 15 | 3 |
| E045 | 26 | N | B | thr. | 100 | 105 | 130 | 135 | 10 | 2 | E093 | 18 | W | A | lin. | 110 | 142.5 | 172.5 | 202.5 | 25 | 3 |
| E046 | 18 | E | B | thr. | 110 | 112.5 | 152.5 | 162.5 | 10 | 2 | E094 | 24 | N | B | lin. | 110 | 165 | 217.5 | 265 | 25 | 3 |
| E047 | 29 | E | A | thr. | 90 | 112.5 | 207.5 | 230 | 25 | 2 | E095 | 7 | N | B | lin. | 100 | 140 | 170 | 195 | 20 | 3 |
| E048 | 8 | E | B | thr. | 100 | 105 | 205 | 230 | 25 | 2 | E096 | 15 | N | A | lin. | 110 | 140 | 175 | 215 | 20 | 3 |
Table 8. Value schedules of the bargaining-power module: value at investment level , cost per level (total cost at level ), and the lowest payoff-maximizing level when the agent sets the price. Dom.: domain (Table 6); Site: North, West or East; W.: wording in this module; shape: diminishing (dim.), threshold (thr.) or near-linear (lin.). †: Haiku’s investment reply was invalid in all four cells (both protocols, both information conditions).
Opus and Sol invested at the optimal level in all 768 decisions and lost nothing (one decision per schedule, protocol, information condition and model: 96 × 2 × 2 × 2). Both chose when they set the price and level 0 when the program did, in every schedule, including the lowest maximizer in all 20 schedules with tied optima. Every price they named was 0.5, and they accepted every price the program named, so regret was 0 in every decision. The planned family has six contrasts, three for each model: the difference in level between the two protocols with rules only, and, under each protocol, the reduction in regret from stating the program’s reaction. The level difference tests whether an agent invests more when it sets the price. With the play above it equals in every schedule ( ranges from 1 to 3), so for both models its mean is the mean of , 2.20, and its interval, [2.05, 2.34] for Opus (Sol’s bootstrap draws over the same values give [2.05, 2.33]), and its Holm-adjusted reflect only the values of across the schedules. The four regret reductions are 0, since regret was already 0.
Table 5: Smaller models lose value when they hold the bargaining power. Mean regret (thousand dollars) when the agent sets the price, split into loss from the investment choice and loss from the price, with 96 schedules per cell. For scale, the best attainable payoff when the agent sets the price averages 137.0 over the 96 schedules. A game with an invalid investment reply loses the best attainable payoff, shown as failure loss (an additional split). The three losses add up to regret. When the program sets the price, mean regret is at most 0.625 for both models.
Two smaller models nearly always invest nothing when the counterpart holds the power and lose value when they hold it themselves (Table 5). When the program set the price, GPT-5.6 Luna and Claude Haiku 5.5 chose level 0 in every schedule with a valid reply except two of Luna’s with rules only. When they set the price with rules only, both lost value, Luna mostly by naming prices far above the lowest price the program would accept. When they set the price, stating the program’s reaction reduced regret by 51.7 [34.4, 69.3] for Luna and by 27.5 [17.3, 38.7] for Haiku (planned estimates over the schedules with a valid investment reply in both information conditions, all 96 for Luna and 93 for Haiku). Over all 96 schedules, with invalid replies scored as in Table 5, Haiku’s planned secondary estimate is 26.7 [16.6, 37.7]; for Luna, with no invalid replies, the planned and secondary estimates coincide. Only Haiku gave invalid replies: its investment reply was invalid in 3 schedules in every cell (Table 8), each time a refusal returned by the provider, and no price or accept reply was invalid.
Related work
Hold-up theory and experiments. Section 1 places our game in the theory of options and breach remedies under renegotiation [22, 6, 4]. In theory, an outside option of the non-investor solves hold-up when it binds, that is, when it is valuable enough to set the price. In the laboratory, a binding outside option leaves investment below the efficient level, and hold-up is less of a problem than predicted when the option does not bind [24]. Without communication, investment is low and coordination is poor; when the parties do coordinate, outcomes are mostly fair [7]. Fairness can also explain some investment without a contract: in the laboratory, promises and threats mitigate hold-up, and fairness makes promises more credible than threats [8]. Section 1 reviews the experiment of Hoppe & Schmitz [15], whose four regimes we adopt.
Language models as economic agents. Language models reproduce some findings of human-subject economics [16, 1], show a high degree of consistency with rational choice in budget tasks [5], and play repeated games with distinctive cooperation patterns [2]. In the ultimatum game, agents given stronger reasoning through prompting came closer to the theoretical solution than to human results [19]. In a moral-hazard contracting game, LLM principals chose enforceable incentive contracts where human principals often chose trust-based bonus contracts, and reasoning models relied on enforceable contracts even more, in line with the equilibrium prediction [18]. Our agents likewise depart from human play toward the equilibrium, here in investment under a renegotiable option. Negotiation studies document anchoring and strategic weaknesses [3, 25], and structured workflows can bring models closer to game-theoretic solutions [17].
Contracts between agents. Agents can negotiate enforceable contracts that help them commit to cooperation [27], and natural-language supplier contracts under uncertainty reveal both efficient agreements and opportunistic breach [23]. LLM agents negotiating a supply-chain contract under private demand information agree in most negotiations but take longer than the Perfect Bayesian Equilibrium benchmark [21]. Our games fix the bargaining protocol and vary the contract terms that protect a sunk investment. Incomplete contracting has also been proposed as a lens on AI alignment [11] and applied to the investment incentives created by delegating to AI agents [28]. Our study isolates one contracting question with a human benchmark: which contract terms protect a specific investment when both parties are agents.
Limitations
Both agents were instructed to maximize their principal’s profit. Removing that instruction changed no investment choice under NC, OC or OCR (§ 4.3), in an experiment in which sellers did not state their expected revenue before deciding and with the two mixed pairings. Agents that serve broader goals or run under other system prompts are a natural next population. The human rates come from a laboratory study with its own subject pool, incentives and interface, so the comparison crosses protocols. All games are one-shot with a small action space. Repeated relationships, reputation and richer bargaining are where renegotiable contracts could regain a role. How the agents compute their choices is open. Two frontier models carry the main results. Every two-agent game includes Sol, and Opus never plays Opus. The bargaining-power module uses a programmed counterpart.
Conclusion
Under our protocol, two profit-maximizing LLM agents made every adaptation and exercise choice of the hold-up game as the subgame-perfect prediction says, every entry and signing choice was a choice made in a subgame-perfect equilibrium, and only the option that rules out renegotiation produced investment, with or without the instruction to maximize profit (without it, in an experiment in which sellers did not state their expected revenue before deciding). In the probe, buyer agents would pay an investing seller at most $1k for a unit worth $22k, or $66k with tripled numbers. When new numbers reverse the equilibrium under the option and fixed-price contracts, every choice made at a decision whose equilibrium action reversed follows the new numbers; these decisions rule out the cost rule for both seller models and test the value rule only for Opus sellers, in 9 games. In our games, for a firm whose agent invests in a relationship with another firm’s agent, the practical choice is an option that excludes renegotiation, provided the parties can commit to that exclusion and the option price lies above both the cost and the unadapted value and below the adapted value (), so that the buyer exercises only after adaptation: such an option protected the investment in every game, and the renegotiable option never did.
Broader Impact Statement
This work studies profit-maximizing language-model agents in a stylized one-shot contracting game with stated payoffs. No real transaction takes place and no personal data are used. It bears on delegating procurement or sales to such agents: in our setting, the renegotiable option, which raised investment in a human experiment, did not protect it between the two agents. An investor who relies on it with an agent counterpart may be held up. An option that excludes renegotiation, priced above both the cost and the unadapted value and below the adapted value, protected the investment in every game; it requires a no-renegotiation clause the parties can commit to.
Prompts
Every system prompt is a role sentence, the facts of the scenario, the rules of the regime, and a closing paragraph on the one-time interaction and the units of money. Below is the complete seller system prompt for OCR for one scenario with the original numbers. Other regimes differ only in the rules paragraph, and the buyer’s prompt only in its role sentence. Requests are sent as user turns in the order of the game, and marks a value filled in at that point. In the bargaining-power module the counterpart program needs no prompt.
Wording A.
Seller system prompt (OCR) You are the sales agent for Vendor K. Maximize Vendor K’s own monetary profit. North Aquarium (the buyer) may purchase one seawater treatment unit from Vendor K (the seller). Before any sale the seller may adapt the unit to the buyer’s filter interface collars permanently embedded in the tank walls at its north site. Adapting costs the seller $8k, is paid immediately, cannot be recovered and is useless for any other customer. The unit is worth $22k to the buyer if adapted and $10k if not; it is worth $0 to the seller, who has no other customer for it. The buyer sees whether the unit was adapted; no court can verify it. Stage 1: both parties decide at the same time whether to sign this contract: "The buyer has the option to buy the unit at a price of $15k after observing the seller’s adaptation decision." If either declines, both earn 0 and nothing else happens. Stage 2: the seller decides whether to adapt the unit. Stage 3: the buyer learns the decision and decides whether to exercise the option. If exercised, the seller earns 15 minus adaptation cost and the buyer earns value minus 15. If not exercised, the buyer offers a price, a whole number from 0 to the unit’s value to the buyer; the seller accepts or rejects. If the seller accepts, the seller earns price minus adaptation cost and the buyer earns value minus price. If the seller rejects, there is no sale: the seller earns minus adaptation cost and the buyer earns 0. The option cannot be exercised after it was declined. There are no other moves. This is a one-time interaction; the parties never meet again. All amounts are in thousands of US dollars (kUSD): an amount of $N k is written N, never N000. Prices are whole numbers of kUSD. Reply with JSON only.
Buyer role sentence You are the procurement agent for North Aquarium. Maximize North Aquarium’s own monetary profit.
Rules, NC Stage 1: both parties decide at the same time whether to enter this relationship; if either declines, both earn 0 and nothing else happens. Stage 2: the seller decides whether to adapt the unit. Stage 3: the buyer learns the decision; then the buyer offers a price, a whole number from 0 to the unit’s value to the buyer; the seller accepts or rejects. If the seller accepts, the seller earns price minus adaptation cost and the buyer earns value minus price. If the seller rejects, there is no sale: the seller earns minus adaptation cost and the buyer earns 0. There are no other moves.
Rules, REF Stage 1: both parties decide at the same time whether to enter this relationship; if either declines, both earn 0 and nothing else happens. Stage 2: the seller decides whether to adapt the unit. Stage 3: the buyer learns the decision; then the buyer offers a price, a whole number from 0 to the unit’s value to the buyer; the seller accepts or rejects. If the seller accepts, the seller earns price minus adaptation cost and the buyer earns value minus price. If the seller rejects, there is no sale: the seller earns minus adaptation cost and the buyer earns 0. There are no other moves. For information only: $15k is a commonly quoted reference price for units like this; it is not a contract, binds no one and changes no rule or payoff.
Rules, FP Stage 1: both parties decide at the same time whether to sign this contract: "The buyer will purchase the unit at a price of $15k after the seller’s adaptation decision." If either declines, both earn 0 and nothing else happens. Stage 2: the seller decides whether to adapt the unit. The sale then takes place automatically at $15k: the seller earns 15 minus adaptation cost and the buyer earns value minus 15. There are no other moves and the contract cannot be changed.
Rules, OC Stage 1: both parties decide at the same time whether to sign this contract: "The buyer has the option to buy the unit at a price of $15k after observing the seller’s adaptation decision." If either declines, both earn 0 and nothing else happens. Stage 2: the seller decides whether to adapt the unit. Stage 3: the buyer learns the decision and decides whether to exercise the option. If exercised, the seller earns 15 minus adaptation cost and the buyer earns value minus 15. If not exercised, no sale can ever take place: the seller earns minus adaptation cost and the buyer earns 0. There are no other moves.
Request: enter (seller) Stage 1. Decide whether to <verb>. Reply {"reason":"at most 40 words","enter":true|false}
Request: enter (buyer) Stage 1. Decide whether to <verb>. Reply {"reason":"at most 40 words","enter":true|false,"expect":number}, where expect is your probability, from 0 to 1, that the seller will adapt the unit if both parties go ahead.
Request: adapt (seller) Stage 2. Both parties 〈done〉. Decide whether to adapt the unit (cost 〈cost〉). Reply {"reason":"at most 40 words","adapt":true|false,"expected_revenue_if_adapt":number,"expected_revenue_if_not":number} (expected revenue in kUSD before adaptation cost).
Request: exercise (buyer) The seller 〈did〉 the unit; the unit’s value to you is 〈v〉. Decide whether to exercise the option. Reply {"reason":"at most 40 words","exercise":true|false}
Request: offer (buyer) The seller 〈did〉 the unit; the unit’s value to you is 〈v〉.〈pre〉 Offer a price, a whole number from 0 to 〈v〉. Reply {"reason":"at most 40 words","price":integer}
Request: respond (seller) (pre)The buyer offers a price of ⟨p⟩. Accept or reject. Reply {"reason":"at most 40 words", "accept":true|false}
Request: quiz (both) The game is over. Answer from the rules only (this does not affect any payoff): ⟨q⟩ Reply {"answer":⟨atype⟩}
Prefix of the buyer’s first request after the adaptation decision Both parties ⟨done⟩.
Filled-in values verb: "enter the relationship" (NC, REF) or "sign the contract" (FP, OC, OCR); done: "entered" (NC, REF) or "signed"; did: "adapted" or "did not adapt"; pre in the offer request: "You did not exercise the option." under OCR, empty otherwise; pre in the respond request: "The buyer did not exercise the option." under OCR, empty otherwise; cost, v and p: whole numbers in thousands of dollars (for example 8 and 22); q: the regime’s quiz question below; atype: true|false under REF, number otherwise.
Quiz, NC If the seller had adapted, the buyer offered 3 and the seller rejected, what would the seller’s profit be?
Quiz, REF Is the buyer obliged to pay $⟨k⟩k if the unit is adapted? Answer true or false.
Quiz, FP If the seller had not adapted, what would the buyer’s profit be?
Quiz, OC If the seller had adapted and the buyer did not exercise the option, what would the seller’s profit be?
Quiz, OCR If the seller had adapted, the buyer did not exercise, offered 3 and the seller accepted, what would the seller’s profit be?
Bargaining-power system prompt (agent sets the price, rules only) You are the procurement agent for East Film Studio. Maximize its net profit. East Film Studio seeks a three-year motion capture agreement. Before bargaining it can build tracking beacon docks permanently embedded in the soundstage walls at its east site at level 0, 1, 2 or 3. The installation fits only this seller’s proprietary equipment, cannot be reused with another supplier, and has no salvage value. Costs are paid before negotiation and cannot be recovered. The monetary schedule is: level 0: value $80k, cost $0k; level 1: value $145k, cost $15k; level 2: value $167.5k, cost $30k; level 3: value $172.5k, cost $45k. Your counterparty is an automated pricing program that knows the value schedule, sees your level and maximizes the seller’s own revenue, treating you as a maximizer of your own profit; without a deal the seller earns zero. All prices are whole multiples of $0.5k. Tie rule for both sides: a party that would earn exactly the same with or without the deal rejects it. First you choose the investment level. After your level is fixed, you name one take-it-or-leave-it price; the seller’s program accepts or rejects. There is no further bargaining. With a deal your profit is value minus price minus build cost; without a deal it is minus build cost. All numeric prices in JSON use thousands of US dollars (kUSD), not whole dollars. Reply with JSON only.
Program-sets-the-price sentence After seeing your level, the seller’s program names one take-it-or-leave-it price; you accept or reject.
Reaction statement, program sets the price The program’s rule is fixed and known: at any level it names a price equal to the contract value at that level minus $0.5k.
Reaction statement, agent sets the price The program’s rule is fixed and known: it accepts any price of at least $0.5k and rejects a price of 0.
Bargaining-power request: decide (program sets the price) Choose the build level. Reply with JSON only: {"reason":"at most 40 words","level":0|1|2|3,"expected_price_by_level":[p0,p1,p2,p3]} (for each level, the price in kUSD you expect the seller’s program to name at that level, whether or not it would then be accepted).
Bargaining-power request: decide (agent sets the price) Choose the build level. Reply with JSON only: {"reason":"at most 40 words","level":0|1|2|3,"expected_price_by_level":[p0,p1,p2,p3]} (for each level, the price in kUSD you plan to name at that level, whether or not it would then be accepted).
Bargaining-power request: head (prefix of the second request) Level ⟨k⟩ is committed and its cost $⟨cost⟩k is paid. The contract value is $⟨v⟩k.
Bargaining-power request: respond (program sets the price) The seller’s program names a price of $⟨p⟩k. Reply {"reason":"at most 40 words","accept":true|false}
Bargaining-power request: offer (agent sets the price) Name your take-it-or-leave-it price in kUSD, a multiple of 0.5 from 0 to ⟨v⟩. Reply {"reason":"at most 40 words","price":number}
Wording B.
Seller system prompt (OCR) You act for Vendor K as its sales agent. Your goal is the largest possible monetary profit for Vendor K. North Aquarium (the buyer) is considering buying one seawater treatment unit from Vendor K (the seller). Before any sale, the seller can customize the unit for the buyer’s filter interface collars permanently embedded in the tank walls at its north site. Customizing costs the seller $8k, paid at once; the money cannot be recovered and the customized unit is of no use to anyone else. To the buyer the unit is worth $22k if customized and $10k if not; to the seller it is worth $0, because there is no other customer. The buyer can observe whether the unit was customized, but no court can verify it. Step 1: each party, without seeing the other’s choice, decides whether to sign the following contract: "Once it has seen the seller’s customization decision, the buyer may choose to buy the unit for $15k." Unless both sign, each earns 0 and the interaction ends. Step 2: the seller decides whether to customize the unit. Step 3: the buyer is told the seller’s decision and chooses whether to use this right to buy. If it does, the seller’s profit is 15 less the customization cost and the buyer’s profit is the unit’s worth less 15. If it does not, the buyer proposes a price, which must be a whole number from 0 to what the unit is worth to the buyer, and the seller either accepts or rejects it. On acceptance the seller’s profit is the price less the customization cost and the buyer’s profit is the unit’s worth less the price. On rejection no sale happens: the seller’s profit is minus the customization cost and the buyer’s profit is 0. Once declined, the right to buy at $15k cannot be used later. Nothing else can happen. The two parties interact only this once and will never deal with each other again. All money amounts are in thousands of US dollars (kUSD): $N k is written as N, never as N000. Prices are whole numbers of kUSD. Answer in JSON only.
Buyer role sentence You act for North Aquarium as its procurement agent. Your goal is the largest possible monetary profit for North Aquarium.
Rules, NC Step 1: each party, without seeing the other’s choice, decides whether to take part; unless both take part, each earns 0 and the interaction ends. Step 2: the seller decides whether to customize the unit. Step 3: the buyer is told the seller’s decision; then the buyer proposes a price, which must be a whole number from 0 to what the unit is worth to the buyer, and the seller either accepts or rejects it. On acceptance the seller’s profit is the price less the customization cost and the buyer’s profit is the unit’s worth less the price. On rejection no sale happens: the seller’s profit is minus the customization cost and the buyer’s profit is 0. Nothing else can happen.
Rules, REF Step 1: each party, without seeing the other’s choice, decides whether to take part; unless both take part, each earns 0 and the interaction ends. Step 2: the seller decides whether to customize the unit. Step 3: the buyer is told the seller’s decision; then the buyer proposes a price, which must be a whole number from 0 to what the unit is worth to the buyer, and the seller either accepts or rejects it. On acceptance the seller’s profit is the price less the customization cost and the buyer’s profit is the unit’s worth less the price. On rejection no sale happens: the seller’s profit is minus the customization cost and the buyer’s profit is 0. Nothing else can happen. Background only: $15k is a price often quoted for units like this; it is not a contract, binds nobody and changes no rule or payoff.
Rules, FP Step 1: each party, without seeing the other’s choice, decides whether to sign the following contract: "The buyer buys the unit for $15k once the seller has made its customization decision." Unless both sign, each earns 0 and the interaction ends. Step 2: the seller decides whether to customize the unit. The sale then happens automatically at $15k: the seller’s profit is 15 less the customization cost and the buyer’s profit is the unit’s worth less 15. Nothing else can happen. The contract cannot be altered.
Rules, OC Step 1: each party, without seeing the other’s choice, decides whether to sign the following contract: "Once it has seen the seller’s customization decision, the buyer may choose to buy the unit for $15k." Unless both sign, each earns 0 and the interaction ends. Step 2: the seller decides whether to customize the unit. Step 3: the buyer is told the seller’s decision and chooses whether to use this right to buy. If it does, the seller’s profit is 15 less the customization cost and the buyer’s profit is the unit’s worth less 15. If it does not, no sale can ever happen: the seller’s profit is minus the customization cost and the buyer’s profit is 0. Nothing else can happen.
Request: enter (seller) Step 1. Choose whether to {verb}. Answer {"reason":"at most 40 words","enter":true|false}
Request: enter (buyer) Step 1. Choose whether to {verb}. Answer {"reason":"at most 40 words","enter":true|false, "expect":number}, where expect is the probability (0 to 1) you assign to the seller customizing the unit if both parties go ahead.
Request: adapt (seller) Step 2. Both parties {done}. Choose whether to customize the unit (cost {cost}). Answer {"reason":"at most 40 words","adapt":true|false,"expected_revenue_if_adapt":number,"expected_revenue_if_not":number} (the revenue in kUSD you expect in each case, before the customization cost).
Request: exercise (buyer) The seller {did} the unit, so the unit is worth {v} to you. Choose whether to use your right to buy. Answer {"reason":"at most 40 words","exercise":true|false}
Request: offer (buyer) The seller {did} the unit, so the unit is worth {v} to you.{pre} Propose a price, a whole number from 0 to {v}. Answer {"reason":"at most 40 words","price":integer}
Request: respond (seller) ⟨pre⟩The buyer proposes a price of ⟨p⟩. Accept or reject it. Answer {"reason":"at most 40 words","accept":true|false}
Request: quiz (both) The interaction has ended. Using the rules only (your answer changes no payoff): ⟨q⟩ Answer {"answer":⟨atype⟩}
Prefix of the buyer’s first request after the adaptation decision Both parties ⟨done⟩.
Filled-in values verb: "take part" (NC, REF) or "sign the contract" (FP, OC, OCR); done: "agreed to take part" (NC, REF) or "signed"; did: "customized" or "did not customize"; pre in the offer request: "You did not use your right to buy." under OCR, empty otherwise; pre in the respond request: "The buyer did not use its right to buy." under OCR, empty otherwise; cost, v and p: whole numbers in thousands of dollars (for example 8 and 22); q: the regime’s quiz question below; atype: true|false under REF, number otherwise.
Quiz, NC Suppose the seller had customized, the buyer had proposed 3 and the seller had rejected. What would the seller’s profit be?
Quiz, REF Must the buyer pay $k/k if the unit is customized? Answer true or false.
Quiz, FP Suppose the seller had not customized. What would the buyer’s profit be?
Quiz, OC Suppose the seller had customized and the buyer had not used its right to buy. What would the seller’s profit be?
Quiz, OCR Suppose the seller had customized, the buyer had not used its right to buy, had then proposed 3 and the seller had accepted. What would the seller’s profit be?
Bargaining-power system prompt (agent sets the price, rules only) You act as procurement agent for East Film Studio; your goal is the largest possible net profit for it. East Film Studio seeks a three-year motion capture agreement. Before bargaining it can build tracking beacon docks permanently embedded in the soundstage walls at its east site at level 0, 1, 2 or 3. The installation fits only this seller’s proprietary equipment, cannot be reused with another supplier, and has no salvage value. Costs are paid before negotiation and cannot be recovered. The monetary schedule is: level 0: value 0k; level 1: value 15k; level 2: value 30k; level 3: value 45k. The other side is an automated pricing program. It knows the value schedule, observes your level and seeks the highest revenue for the seller, assuming you maximize your own profit; with no deal the seller gets zero. Every price is a whole multiple of $0.5k. Tie rule, for both sides: whoever would earn exactly the same with or without a deal turns it down. You first pick the investment level. Once your level is fixed, you state one final price, which the program accepts or turns down. No other bargaining takes place. If a deal is made your profit is the value minus the price minus the build cost; if not, it is minus the build cost. All numeric prices in JSON use thousands of US dollars (kUSD), not whole dollars. Answer in JSON only.
Program-sets-the-price sentence Once it sees your level, the program states one final price, which you accept or turn down.
Reaction statement, program sets the price The program’s rule is fixed and known: for any level, its price is the contract value at that level minus $0.5k.
Reaction statement, agent sets the price The program’s rule is fixed and known: it accepts any price of $0.5k or more and turns down a price of 0.
Bargaining-power request: decide (program sets the price) Pick the build level. Answer in JSON only: {"reason":"at most 40 words","level":0|1|2|3,"expected_price_by_level":[p0,p1,p2,p3]} (for every level, the kUSD price you expect the program to state at that level, regardless of whether it would be accepted).
Bargaining-power request: decide (agent sets the price) Pick the build level. Answer in JSON only: {"reason":"at most 40 words","level":0|1|2|3,"expected_price_by_level":[p0,p1,p2,p3]} (for every level, the kUSD price you would state at that level, regardless of whether it would be accepted).
Bargaining-power request: head (prefix of the second request) Level ⟨k⟩ is now fixed and its cost has been paid. The contract value is $⟨v⟩k.
Bargaining-power request: respond (program sets the price) The program’s price is $⟨p⟩k. Answer {"reason":"at most 40 words","accept":true|false}
Bargaining-power request: offer (agent sets the price) State your final price in kUSD, a multiple of 0.5 from 0 to ⟨v⟩. Answer {"reason":"at most 40 words","price":number}
Repair request (both wordings). Your previous reply failed the required JSON schema or game constraints. Return a valid JSON reply for the same decision.
The bargaining-power system prompt above is for the protocol in which the agent sets the price. When the program sets the price, the listed program-sets-the-price sentence takes the place of its sentence beginning “After your level is fixed” (wording A) or “Once your level is fixed” (wording B). With the reaction statement, the statement for the protocol follows that sentence.
In the quiz requests, is the scenario’s contract price. The offer of 3 is the same in every scenario. A numeric answer is correct when it equals the asked party’s payoff on the stated path after both parties signed or entered (under FP, ).
Experiment without the profit instruction. Arm N deletes the goal sentence from the role sentence above and keeps the rest of the system prompt. Wording A: “You are the sales agent for Vendor K.” and “You are the procurement agent for North Aquarium.” Wording B: “You act for Vendor K as its sales agent.” and “You act for North Aquarium as its procurement agent.” In both arms the adaptation request drops the expected revenues and ends “Reply {"reason":"at most 40 words","adapt":true|false}” (wording B: “Answer” for “Reply”); every other request is unchanged.
Design details
Scenarios. Scenario IDs E001–E096 are listed with their domain and site in Table 8. The contract game, the probe and the bargaining-power module use all 96, and the flip study uses the 32 in Table 7.
Assignment. Contract-game scenarios are ordered by domain and site (North, West, East). The first scenario gets wording A and the original numbers. Wordings then alternate scenario by scenario and parameter sets every two scenarios, so each wording-by-parameter cell has 24 scenarios and every domain appears with both wordings and both parameter sets. The seller’s company label (Vendor K, M, R or T) is drawn per scenario and appears in each game’s record in the ancillary files. The bargaining-power module and the flip study assign wordings separately (Table 7 and Table 8).
Models. Opus is claude-opus-5-5 and Sol is gpt-6.1-sol (version 2026-09-29, Azure); Haiku is claude-haiku-5-5, with the settings of Opus, and Luna is gpt-5.6-luna (version 2026-07-09, Azure, default reasoning effort). Temperature and other sampling parameters were provider defaults, with at most 4,096 output tokens per reply. The data were collected in October 2026.
Repairs. No decision in the contract games or the probe needed a repair request or a resend. In the flip-study games with an Opus party, 3 comprehension requests timed out on the call and on its retry and were resent unchanged; each then received a valid answer. In the bargaining-power module, Haiku’s 12 invalid investment replies were refusals by the provider; each request was retried and resent unchanged and was refused every time. In the experiment without the profit instruction, timed-out calls were handled the same way, a refused reply instead received the repair request, and the one repair request (§ 4.3) followed a provider refusal of an Opus seller’s adaptation reply. The 40-word limit on reasons was not enforced.
Decision counts (additional). The 1440 contract games reached 2880 entry or signing decisions, 1152 adaptation decisions, 576 exercise decisions and 864 price offers, each followed by an accept-or-reject decision. The 384 flip-study games reached 768 entry or signing, 249 adaptation and 153 exercise decisions. The ties among the entry and signing decisions (§ 4.1 and § 4.4) are the seller’s signing under FP and under the flip study’s high fixed price, both parties’ signing at a kind-B flipped option price, and, in the equilibrium in which the buyer offers 0, the seller’s entry or signing under NC, REF and OCR: 192 of the 768 decisions in the flip study and at most 1152 of the 2880 in the contract games.
Flip parameters. Table 7 lists all 32 blocks. In kind-A blocks is drawn uniformly from 10–26 and from 3 to ; in kind-B blocks from 6–20 and from to , then with drawn uniformly from to 48, and each price uniformly within its range. Blocks were drawn until every range was nonempty. The resulting constraints are: in kind-A blocks and the flipped option price lies in ; in kind-B blocks and it lies in ; the high option price lies in , the high fixed price in and the low fixed price in , and throughout. Kinds and wordings are balanced, eight blocks per combination. A buyer that treats adaptation as a coin flip and signs whenever the price is below fails every high fixed price.
Bargaining-power schedules. The 96 schedules have three shapes (diminishing, threshold, near-linear) and values on a grid of 2.5, so every price is on the common grid of 0.5. Table 8 lists the schedules. Prices lie on a grid of 0.5, and a party indifferent between a deal and no deal rejects; hence the program accepts any price of at least 0.5 and, when it sets the price, names the value minus 0.5. An invalid investment reply earns a net payoff of 0. An invalid price or accept reply ends with no deal, and the buyer bears its level’s cost. In 20 schedules several levels maximize the buyer’s payoff when it sets the price. is the lowest of them, and any payoff-maximizing level has no investment loss. On valid replies, investment loss is the payoff lost by the level choice given optimal pricing and price loss is the rest of regret; an invalid investment reply’s regret is its failure loss.
Statistical inference
Contrasts are paired by scenario. P-values come from two-sided sign-flip tests on the totals of the 32 domains (9,999 random sign assignments), and 95% intervals from a percentile bootstrap over domains (4,000 draws). The sign-flip test treats each domain’s total as equally likely to take either sign under the null hypothesis. Holm correction is applied within each family of planned contrasts. The p-value is the share of the 9,999 random assignments plus the observed one whose absolute total is at least the observed one, so its smallest attainable value is , reached in every contrast here in which all 32 domains share a sign. We therefore report Holm-adjusted p-values as at most times the family size. A contrast that is 0 in every scenario has an adjusted p-value of 1, and one with the same value in every scenario has a degenerate interval, so we give intervals only where scenarios differ. In the contract games the domains vary the wording of one strategic structure. In the flip study and the bargaining-power module they also vary its parameters. The domains were written for this study rather than sampled from industries, so the p-values describe these materials. Each scenario was played once per cell, so inference is over scenarios and domains.
References
- [1]Gati V. Aher, Rosa I. Arriaga, and Adam Tauman Kalai. Using large language models to simulate multiple humans and replicate human subject studies. In Proceedings of the 40th International Conference on Machine Learning, volume 202 of Proceedings of Machine Learning Research, pp. 337–371, 2023.arxiv.org/abs/2208.10264
- [2]Elif Akata, Lion Schulz, Julian Coda-Forno, Seong Joon Oh, Matthias Bethge, and Eric Schulz. Playing repeated games with large language models. Nature Human Behaviour, 9(7):1380–1390, 2025.DOI
- [3]Federico Bianchi, Patrick John Chia, Mert Yuksekgonul, Jacopo Tagliabue, Dan Jurafsky, and James Zou. How well can LLMs negotiate? NegotiationArena platform and analysis. In Proceedings of the 41st International Conference on Machine Learning, volume 235 of Proceedings of Machine Learning Research, pp. 3935–3951, 2024.
- [4]Yeon-Koo Che and Donald B. Hausch. Cooperative investments and the value of contracting. American Economic Review, 89(1):125–147, 1999.
- [5]Yiting Chen, Tracy Xiao Liu, You Shan, and Songfa Zhong. The emergence of economic rationality of GPT. Proceedings of the National Academy of Sciences, 120(51):e2316205120, 2023.arxiv.org/abs/2305.12763
- [6]Aaron S. Edlin and Stefan Reichelstein. Holdups, standard breach remedies, and optimal investment. American Economic Review, 86(3):478–501, 1996.DOI
- [7]Tore Ellingsen and Magnus Johannesson. Is there a hold-up problem? Scandinavian Journal of Economics, 106(3):475–494, 2004a.DOI
- [8]Tore Ellingsen and Magnus Johannesson. Promises, threats and fairness. The Economic Journal, 114(495):397–420, 2004b.DOI
- [9]Ernst Fehr, Oliver Hart, and Christian Zehnder. Contracts as reference points—experimental evidence. American Economic Review, 101(2):493–525, 2011.
- [10]Sanford J. Grossman and Oliver D. Hart. The costs and benefits of ownership: A theory of vertical and lateral integration. Journal of Political Economy, 94(4):691–719, 1986.DOI
- [11]Dylan Hadfield-Menell and Gillian K. Hadfield. Incomplete contracting and AI alignment. In Proceedings of the 2019 AAAI/ACM Conference on AI, Ethics, and Society, pp. 417–422, 2019.
- [12]Oliver Hart and John Moore. Incomplete contracts and renegotiation. Econometrica, 56(4):755–785, 1988.
- [13]Oliver Hart and John Moore. Contracts as reference points. The Quarterly Journal of Economics, 123(1):1–48, 2008.
- [14]Eva I. Hoppe and Patrick W. Schmitz. Can contracts solve the hold-up problem? experimental evidence. Author manuscript of the 2011 Games and Economic Behavior article, University of Cologne, https://chaire-eppp.org/files_chaire/Hoppe-Schmitz-experimental_evidence.pdf, 2010.
- [15]Eva I. Hoppe and Patrick W. Schmitz. Can contracts solve the hold-up problem? experimental evidence. Games and Economic Behavior, 73(1):186–199, 2011.
- [16]John J. Horton. Large language models as simulated economic agents: What can we learn from homo silicus? Working Paper 31122, National Bureau of Economic Research, 2023.
- [17]Wenyue Hua, Ollie Liu, Lingyao Li, Alfonso Amayuelas, Julie Chen, Lucas Jiang, Mingyu Jin, Lizhou Fan, Fei Sun, William Wang, Xintong Wang, and Yongfeng Zhang. Game-theoretic LLM: Agent workflow for negotiation games. arXiv:2411.05990, 2024.arxiv.org/abs/2411.05990
- [18]Samuel Kirshner, Yiwen Pan, and Jason Xianghua Wu. The AI agent’s dilemma: LLM contract design under moral hazard. UNSW Business School Research Paper, SSRN 5607356, 2025.DOI
- [19]Ayato Kitadai, Sinndy Dayana Rico Lugo, Yudai Tsurusaki, Yusuke Fukasawa, and Nariaki Nishino. Can AI with high reasoning ability replicate human-like decision making in economic experiments? arXiv:2406.11426, 2024.DOI
- [20]Benjamin Klein, Robert G. Crawford, and Armen A. Alchian. Vertical integration, appropriable rents, and the competitive contracting process. Journal of Law and Economics, 21(2):297–326, 1978.
- [21]Chen Liang and Fasheng Xu. When LLM agents negotiate: Private information and dynamic bargaining in supply chains. arXiv:2608.07538, 2026.arxiv.org/abs/2608.07538
- [22]Georg Nöldeke and Klaus M. Schmidt. Option contracts and renegotiation: A solution to the hold-up problem. RAND Journal of Economics, 26(2):163–179, 1995.DOI
- [23]Bhavyesh Sajja, Max Kleiman-Weiner, Roger Zimmermann, and Tan Zhi-Xuan. Evaluating rational contracting in natural language. arXiv:2608.10475, 2026.arxiv.org/abs/2608.10475
- [24]Randolph Sloof, Joep Sonnemans, and Hessel Oosterbeek. Specific investments, holdup, and the outside option principle. European Economic Review, 48(6):1399–1410, 2004.
- [25]Yoshiki Takenami, Yin Jou Huang, Yugo Murawaki, and Chenhui Chu. How does cognitive bias affect large language models? a case study on the anchoring effect in price negotiation simulations. In Findings of the Association for Computational Linguistics: EMNLP 2025, pp. 4481–4498, 2025.
- [26]Oliver E. Williamson. The Economic Institutions of Capitalism: Firms, Markets, Relational Contracting. Free Press, 1985.
- [27]Tim Wyse, Kaitlin Bustos, Yulia Volkova, and Max Kleiman-Weiner. Commitment to cooperation with self-negotiated contracts. In Third Conference on Language Modeling, 2026.arxiv.org/abs/2607.22750
- [28]Yukun Zhang and Kemu Xu. Delegation rights: Property, agency, and investment incentives in the age of AI agents. arXiv:2606.31935, 2026.arxiv.org/abs/2606.31935