betaFeedback

WorksComputer Science

A Self-Evolving Agent Runs Like a Gas: Describing and Predicting Its Evolution with Fluid Dynamics

Abstract

A self-evolving LLM agent improves in service by rewriting its own plugin library under a validation gate. Agent engineering is converging on exactly such agents: single systems in continuous operation, carrying populations of runtime instances (plugins, skills, tools) past the point where anyone can read the library and still predict what the system will do. Yet the run itself, the aggregate motion of that population across hundreds of edit rounds, is unstudied: no theory says whether it relaxes, oscillates, or freezes. A recent kinetic theory supplies the missing level: under four architectural properties, the plugin instances of one agent behave as identical hard spheres whose density obeys a master equation, and the theory is candid that its gate injects and drains credit, momentum is not conserved, and waves are excluded by construction. We argue the boundary is architectural; keeping the four safety theorems, we stack conservation axioms on top. A closed-credit ledger makes credit a collision invariant and unlocks acoustic waves at a measurable, engineerable speed; a zero-sum projection gate conserves centroid momentum unconditionally, for any proposal stream. Both layers are proven and executed: the axis total is conserved to the digit over 66 interventions under a live gate, and the pre-registered decisive probe returns the theory's directional verdict: linear wavefronts where the ledger holds, dissipation where it does not, replicated across two population sizes. For agents that must never stop, the gas description is not a metaphor: it is the measurement interface through which their future becomes readable.

ACM Reference Format:

Haipeng Wang. 2026. A Self-Evolving Agent Runs Like a Gas: Describing and Predicting Its Evolution with Fluid Dynamics: Blue Sky Ideas Track. In Preprint: submitted to AAMAS 2027 (Blue Sky Ideas Track), arXiv, 2026, IFAAMAS, 5 pages.

Introduction

The destination of current agent engineering is an operating system. Academic prototypes already embed LLM agents into OS-style architectures [12]; industrial harnesses converge on everything-is-a-plugin, whose engineering point is hot-swappable capability: tools mount, unmount, and update while the operation keeps running. The trend has consolidated money: Alibaba Cloud’s September 2026 Qwen Book “agent computer,” stated principle OS-as-harness [13]. The empirical caution agrees: split capability across many cooperating agents and coordination becomes the dominant failure mode; multi-agent pipelines frequently lose to a single agent with good tools [2, 4]. Four lines (academic prototypes, industrial consolidation, failure evidence, engineering practice) converge on one design point: one agent in continuous operation, a population of runtime instances growing inside it (MCP servers, plugins, skills, tools, ...). The complexity to master is not a society of agents; it is the population inside one agent. Past a few hundred instances, no one reads the whole library and still understands the system: the run becomes unpredictable, because instances interact and compose in combinations nobody can enumerate.

DeepSeek, born inside quantitative trading where the operation cannot pause for a recompile, shipped the DeepSeek Harness (DSH), a discipline where everything is a plugin: every runtime instance written, composable, auditable, and swappable [5]. Under that discipline self-evolving agents become trainable, the loop well studied: proposing edits, keeping it safe, pricing acceptances [19, 21, 22, 25]. What has not been studied is the agent’s run itself, the aggregate motion of the trainable state across hundreds of edit rounds. Does the running system obey any law: does it relax, oscillate, freeze? Can its endpoint be predicted without running it? The literature has no answer: no level of description at which those questions make sense. The agent as a unit is too coarse: the interesting motion is inside it.

The missing level turned out to be physics. In studying DSH, a recent kinetic theory [20] observed that the 2026 Fields Medal won by Deng (deriving the Boltzmann equation from deterministic hard-sphere dynamics, the mathematical heart of Hilbert’s sixth problem [6, 10]) is exactly the needed move: from microscopic rules to a macroscopic equation of motion. That theory is candid that its fluid layer is locked: no conserved quantity survives its operators, so waves, vortices, and cascades are excluded by construction. We read that diagnosis as a design menu, one constraint up front: the four safety properties are the load-bearing foundation of hot-swap safety, proven theorems, and this program does not touch them. Phenomena are not seen or missed; they are admitted or excluded by axioms, so the program adds axioms rather than removing them: stack conservation laws on top of the untouched foundation, and inherit each new layer’s fluid regime with its equations, its phenomena, and its instruments.

This work is licensed under a Creative Commons Attribution International 4.0 License.

Preprint: submitted to AAMAS 2027 (Blue Sky Ideas Track), Under review, 2026, arXiv. © 2026 Copyright held by the owner/author(s).

The program, stated in one sentence: keep the safety axioms A1–A4 untouched, stack conservation axioms on top of them, derive what fluid phenomena each stacked layer admits, and use the inherited theory of that layer to describe and predict the DSH-plugin agent’s run. The paper proceeds in five steps: the gas the four axioms give and why its fluid layer stays locked (3.1), the stack we add (3.2), what it predicts (3.3), the verification (Section 4), the outlook (Section 5). Our contributions:

  • A conservation stack for plugin agents. Two new axioms (a closed-credit ledger and a zero-sum projection gate) layered on the untouched safety foundation: proofs that the ledger makes credit a collision invariant (a barotropic Euler regime, acoustic waves at engineerable speeds), and that the gate conserves centroid momentum unconditionally, for any proposal stream.

  • Phenomena and instruments with named equations. A forecastable credit-shock failure mode, collective ordering, and six instruments from statistical physics that read population statistics any runtime logs, composed into a differential diagnosis for stalled training.

  • Executed verification. Both laws execute through a live gate to the digit (66 interventions under free, non-conserving proposals), and the pre-registered decisive probe returns the directional verdict (wavefronts where the ledger holds, dissipation where credit washes), replicated across two population sizes.

Related Work

Mean-field and kinetic approaches to multiagent learning fix the architecture and study the dynamics [11, 15, 23]; self-evolution research fixes a loop and studies its safety or gains [21, 22]. This program inverts both: the architecture is the free variable that selects the class of dynamics, the axioms its coordinate system. And where multi-agent designs disperse capability, inheriting their coordination failures [4], this program concentrates it inside one agent: the gas of Section 3.1 is one agent’s interior. Agent-OS architectures supply the habitat [12, 13]; the kinetic ladder is classical mathematics [3, 6, 8, 9]; manufactured emergence [1, 7, 17] yields patterns without equations; this program’s lineage laws come with measurable parameters. What is new is the substrate: conservation laws as legislated runtime policy rather than physical law. Axioms, not scale, are the levers.

Theory: The Conservation Stack Program

The program builds on a foundation it did not lay, which is stated first; proofs and per-run artifacts are deferred to the extended version.

Step Zero: The Gas the Four Axioms Give

DSH’s Cordis kernel [18] proves four structural properties of plugins: A1 permutation symmetry (order-independent composition), A2 mount/unmount reversibility (no residue), A3 acyclic dependencies (no instantaneous self-influence), and A4 typed contracts (bounded, short-range interaction). The kinetic theory [20] shows that under exactly these four properties, the plugin-version instances inside one agent behave as a gas of identical hard spheres, writing a master equation over the plugin×version×task grid,

∂tf+v⋅∇xf+∇v⋅(Ff)=Q[f,f]+R[f]+G[f],(1)\partial_t f + \boldsymbol{v}\cdot\nabla_x f + \nabla_v\cdot(Ff) = Q[f,f] + R[f] + G[f], \tag*{(1)}

with collision QQ (co-activation with attribution), reaction RR (mount/unmount), gate GG (selective dissipation), and external field FF (the task distribution). Because the plugin library is the agent’s entire trainable state, this is not an analogy layered on top of the agent: the agent’s run is, at this level of description, this gas’s dynamics. Its headline signatures are already measured under pre-registered criteria [20]: NESS relaxation (not convergence to an optimum), and −1/2-1/2 fluctuation scaling at sufficient edit density (β^=−0.58\hat{\beta}=-0.58 vs. −1/2-1/2; flat at −0.04-0.04 when sparse).

The lock is structural: the classical route from kinetics to hydrodynamics runs through collision invariants, conserved fields that survive the collision operator and close the moment hierarchy at Euler level [8]. The agent’s gas of Eq. (1) has none: the gate injects and drains credit, reward arrives from outside, and the collision rule mints attribution mass. Every candidate invariant is annihilated at the collision layer.

Corollary 3.1 (No transport without conservation). Under the current axioms (open credit), for every first-order moment φ\varphi some operator T∈{Q,R,G,F}T \in\{Q,R,G,F\} has ∫φ(Tf)≠0\int\varphi(Tf) \ne0; hence no conservation law and no closed pp-system (the classical route to finite-speed transport), so within (1) no library moment carries a closed transport equation (the gate adjudicates on scores, not on moments), and open-credit responses must be mean-field or dissipative (eq:3).

Step One: The Conservation Stack

The first rule is restraint: A1–A4 are the kernel-proven foundation of hot-swap safety, kept in force in every runtime below; what the program adds is a second storey, conservation axioms stacked on top: a closed-credit ledger, a zero-sum projection gate, an energy analogue left open. A1–A4 is demolition, not extension; our release experiments (Section 4) quantify the named failure modes: evidence for the foundation.

Layer 1: the closed-credit ledger

Call a runtime closed-credit if reward is minted once and thereafter only transferred among instances, no operator destroys credit, and the gate adjudicates on shares of the conserved total: each clause is an engineering decision, and the theorem below is a build sheet.

Proposition 3.2 (Conditional Euler regime). Let ff solve (1) on a closed-credit runtime, with the bounded-locality collision structure of [20]. Then M:=∫f dx dvM := \int f\,\mathrm{d}x\,\mathrm{d}v is a collision invariant; given the virial constitutive closure below, the moment hierarchy closes at first order in the conserved field as a hyperbolic pp-system, and credit-density perturbations propagate with finite speed c2=∂p/∂ρc^2 = \partial p/\partial\rho.

The constitutive closure is the surface an architect tunes: attribution transfers of second moment κa:=12∫(Δe)2D(Δe)\kappa_a := \frac{1}{2}\int(\Delta e)^2D(\Delta e), played against per-cell co-activation traffic rc∝ρcr_c \propto\rho_c, amount to the Haff-type equation of state p=κaρ2p=\kappa_a\rho^2 [9]; the conserved pair (ρ,m)(\rho,m) with m:=ρ um:=\rho\,u obeys the system ∂tρ+∇x⋅(ρ u)=0\partial_t\rho+\nabla_x\cdot(\rho\,u)=0, ∂t(ρ u)+∇xp(ρ)=ρF\partial_t(\rho\,u)+\nabla_xp(\rho)=\rho F, whose characteristic speeds u±cu\pm c carry the acoustic speed

c2=p′(ρ)=2κaρ>0,(2)c^2=p'(\rho)=2\kappa_a\rho>0, \tag*{(2)}

so hyperbolicity is immediate and the Cauchy problem is well-posed [3]. Because the runtime is discrete with a complete ledger, this balance is exact: cellwise Ec(t+1)−Ec(t)=∑c′(Φc′→c−Φc→c′)E_c(t+1)-E_c(t)=\sum_{c'}(\Phi_{c'\to c}-\Phi_{c\to c'}), no error term, a ledger identity auditable from the training log.

Layer 2: zero-sum attribution, and its strict form

The next rung is momentum: attribution updates are unconstrained, so internal collisions move ∑ivi\sum_i v_i at will. The zero-sum attribution axiom legislates Newton’s third law: every co-activation’s attribution updates sum to zero on a shared momentum axis (ηiai+ηjaj=0\eta_i a_i+\eta_j a_j=0, where θi←θi+ηiai\theta_i\leftarrow\theta_i+\eta_i a_i is the position update), with a birth/death protocol so mounting and unmounting mint no position. On such a runtime the axis total P=∑iθiP=\sum_i\theta_i is a second collision invariant, provided attribution pairs are bundled into single edit sets so the gate cannot split a zero-sum pair.

Legislation has a weakness: it holds only if proposals comply; gases conserve momentum because the collision operator cannot do otherwise. The strict form is an operator:

Projection axiom (A5-v, mechanistic form). The runtime does not trust proposals: every candidate edit set, before evaluation, passes the zero-sum projection: its net axis displacement dd is computed, and the deficit −d-d is charged to a designated reserve instance. Conservation is a property of the application operator, not of the proposals.

Proposition 3.3 (Unconditional centroid conservation). Under the projection gate, for every proposal stream, including streams that violate any axiom, the accepted library satisfies ∑iθi(t+1)=∑iθi(t)\sum_i\theta_i^{(t+1)}=\sum_i\theta_i^{(t)} across every acceptance: the centroid’s coordinate on the momentum axis (not a credit total) is conserved unconditionally. Under the locality and virial assumptions of Proposition 3.2, the joint (ρ,g)(\rho,g) first-order system is hyperbolic with two acoustic branches cρ=2κaρc_\rho=\sqrt{2\kappa a_\rho} and cv=2κvgc_v=\sqrt{2\kappa v_g}.

Concretely, the gate applies P(E)=E∪{(res,−d(E))}\mathcal{P}(E)=E\cup\{(\mathrm{res},-d(E))\} when the reserve charge is admissible, else drops the set; either way d(P(E))=0d(\mathcal{P}(E))=0, and every set the gate evaluates lies in the zero-sum class. The reserve is the wall: it takes the recoil, is named, and is auditable. In one dimension each conserved pair forms a 2×22\times2 pp-system with its own acoustic branch and Riemann invariants, with credit and displacement waves uncoupled at first order. The stack composes: a further layer never breaks a conservation law already in force (stack monotonicity), which licenses the word. Layer 3, the energy analogue, is the open rung (Section 5).

Step Two: Predicted Phenomena and Instruments

“Wave” is operationalized before it is claimed: a perturbation profile whose radius grows linearly in time, as opposed to instant or decaying:

r(t)=c (t−t0),r≡R, instant and global,resp⁡(r)∝e−r/ξ;r(t)=c\,(t-t_0),\qquad r\equiv R,\ \text{instant and global},\qquad\operatorname{resp}(r)\propto e^{-r/\xi};

wave-like (finite speed)                                             mean-field                                                     dissipative

{#eq:3} $r\geq0$ is the response radius (instances responding, zero before onset $t_0$), and the dissipative profile is read on $r\geq0$; Corollary 3.1 confines the locked runtime to the second and third patterns, so the closed-credit prediction is directional; a measurable speed is a settable requirement: damage containment and intervention timing become wave-speed questions.

A forecastable failure mode: the credit shock. A pp-system’s classical theory is not gentle: compression fronts steepen and shocks form in finite time, a gradient blowup akin to the announced Navier–Stokes construction [14]. Translated: if credit density somewhere grows faster than the acoustic flux can carry it away, the stack’s own equation predicts a discontinuity: a credit shock, the avalanche-like concentration practitioner folklore calls collapse or winner-takes-all. We do not claim observability at minimal scale; the theory names the regime (compression against a bounded acoustic flux) and the quantity to watch (∣∇ρ∣|\nabla\rho| against cρc_\rho), a forecastable failure mode before it is ever observed.

Ordering, and the instruments. A second predicted collective is synchrony: Kuramoto-type populations snap from disorder to order at a critical coupling, a jump invisible in any individual’s telemetry, with the order billed to the environment [16]; the correlation length ξ\xi (diverging) and the four-point susceptibility χ4\chi_4 (peaking) are the early-warning gauges. Synchrony sits outside the stack (locking needs dependency cycles, the deliberately unsafe family), but its moral is the paper’s: the collective is measurable where the individual is not. Because conservation laws fix what the macroscopic variables are, the measurement technology is a century old: six instruments read any runtime’s population logs (aging tt, correlation shift ξ\xi, effective temperature TeffT_{\mathrm{eff}}, transport coefficients D,ηD,\eta, dynamical heterogeneity χ4\chi_4, and fluctuation scaling, field-tested [20]). Table 1 composes them into a differential diagnosis: the three causes of a stalled run separate by joint readings. The Blue Sky claim: debugging a stalled agent should look like a lab session: perturb, relax, measure, rather than read telemetry and guess.

ReadingAging α\alphaGate info (δgate\delta_{\mathrm{gate}} vs δind\delta_{\mathrm{ind}})Accept rateCause →\rightarrow prescription
τ∼twα\tau\sim t_w^\alpha, α>0\alpha> 0; χ4\chi_4 high>0> 0both ≈0\approx0variedstructural freezing →\rightarrow anneal
τ\tau flat; fresh tasks restore gains00δgate>0\delta_{\mathrm{gate}} > 0, δind≈0\delta_{\mathrm{ind}} \approx0normalinformation exhaustion →\rightarrow refresh validation
τ\tau flat; gains stay absent everywhere00both ≈0\approx0→0\rightarrow0proposal exhaustion →\rightarrow enrich reflector / raise edit budget

Table 1. Differential diagnosis for stalled training: three plateau causes separate by joint readings; prescriptions differ.

Experiments

All experiments run on the faithful minimal instance (WebShop text mode [24]; 50–200 four-parameter skill plugins; deterministic stub reflector; strict validation gate), protocols frozen before the runs, numbers reproducible from the archived run artifacts.

Foundation stress-tests. Releasing A4 (typed contracts) lets one out-of-domain edit poison the highest-value document: the improvement waterline stops dead (0.2158 vs. 0.2683 matched control). Releasing A3 (acyclicity) lets dependency cycles form between the two highest-value documents: 63 execution aborts vs. zero in control, mean score untouched, since top-3 retrieval routes around the dead document. Both bites are the theory’s named failure channels; the second is the strongest practical argument for instruments: score-class readings are blind to structure-class failures.

Channel calibration. Perturbing the highest-retrieval document against an undisturbed twin: execution-channel disturbances do not propagate at all (exactly zero response, 6 of 6: the channels are decoupled, so damage can be invisible to retrieval telemetry); retrieval-channel disturbances propagate mean-field (full-library response within one round); the conservation channel (moving credit itself) is the one that can carry a wave.

Figure 1. (a) Closed-credit arm, seed 31: the response radius (instances with credit deviation >0.002> 0.002 from the in-process twin, perturbed pair excluded) grows linearly (frozen-criteria fit of Table 2: R2=0.94R^2 = 0.94, 0.41 instances/round over 20 rounds). (b) The open-credit control, same seed: saturates without a sustained wavefront. (c) Retrieval-channel disturbance: full-neighborhood response within one round (mean-field).

Armseedverdictbest R2R^2max radius/48
closed11dissipative—2
closed21dissipative—12
closed31wave-like0.9412
closed41dissipative—13
closed51wave-like0.9114
open (5 seeds): all dissipative, R2R^2 n/a, max radius 7–11

Table 2. The decisive probe (five seeds per arm, twin-controlled, frozen criteria). “Wave-like” = linear radius growth (R2≥0.9R^2 \ge0.9, ≥5\ge5 rounds); the wave seeds ran linear for 20 and 16 rounds at 0.41/0.65 instances/round.

Conservation executed. The momentum layer promises conservation through a live gate for any proposal stream. Legislated (bundled zero-sum proposals): the axis total fixed to the digit in every seed. Mechanistic (free non-zero-sum proposals through the projection gate): the projection fired 66 times, the axis total conserved exactly in every seed, while the unprojected control drifted, at no measurable performance cost. Identical free streams; only the projected arm conserves: Proposition 3.3 verified. Still unobserved: the displacement wave itself (the cvc_v branch), which needs the wave protocol on a projected, closed-credit arm.

The decisive probe. The theory’s directional prediction (waves only where the ledger holds) was pre-registered with frozen criteria (Eq. (3)), then executed: closed-credit vs. an open control whose gate re-allocates 15% of total credit per round, twins forked in-process from the identical pre-disturbance state, conservation audited to machine precision every round. Three findings. First, the directional prediction holds: the closed arm shows linear wavefronts in 2 of 5 seeds (R2=0.94/0.91R^2 = 0.94/0.91 over 20/16 rounds) at 0.41 and 0.65 instances per round, the first reading of cρc_\rho’s discrete counterpart, while the open control shows none (5 of 5 dissipative, exactly Corollary 3.1’s verdict). Second, mean-field is absent here for both economies (0 of 10): the channel taxonomy is complete. Third, the absolute criterion (3 of 5) was not met (2 of 5), and we report that as it stands.

Larger population (N=200N=200). Re-ran at four times the library (10 seeds per arm, one lost to a logging collision; component-aware perturbation: the co-activation graph’s largest component holds 93 of 200): the verdict replicates, 2 of 9 closed vs. 0 of 10 open; mean-field again absent (0 of 19); normalized reach larger on the closed arm (0.71 vs. 0.60). The batch also corrects the small-instance reading: the binding constraint is signal attenuation, because rewards split over four times the accounts dilute the disturbance within a few hops, so scaling propagation is a ledger-density lever, not a theory limitation. Two pre-registered side measurements returned null: the φ\sqrt{\varphi} speed scaling fell below the noise floor; relaxation time was flat in waiting time, the driven steady state, not aging. Scope: one task domain, a stub reflector; qualitative signatures carry by construction, constants are device-local; a richer reflector could reach regimes the stub cannot.

Outlook and Open Problems

The blue sky. Take self-evolving agents’ trajectory seriously: recursive self-improvement over a plugin library that is the entire trainable state [5]. Its endpoint: an operating-system-like agent, millions of versioned plugins trading credit under gates. At those scales this paper’s questions become operating requirements: how fast does damage travel? is the population aging? which conservation laws held today? And a harder constraint: a downtime-intolerant agent (the trading floor, where the book cannot pause to be debugged) admits no offline inspection; its future must be read off the live run: the gas description’s readouts, as statistical mechanics made Avogadro-scale populations governable. For agent populations that science does not exist yet; this paper builds its first two rungs. For the practitioner this is already an operations manual: install the rules, read the dashboard, diagnose the stall (Table 1), watch the collapse gauge; a one-page quick-start is in the extended version.

Open problems. (1) Axiom independence: is each of A1–A4 independent in the kernel’s theorem graph [18]? (2) The energy analogue: which plugin quantity plays energy (Noether-style: pair kernel symmetries with candidates)? (3) Architecture contracts: specify a blast radius as a wave-speed requirement. (4) Heterogeneous populations: mixed-gas physics for mixed axiom sheets, and instruments as standards.

Conclusion

Four structural axioms turn the agent’s interior into a gas: signatures measurable, failure channels nameable, a conservation stack unlocking a fluid regime whose directional prediction replicates across two scales. The gas description is not a metaphor laid over the agent; it is a measurement interface for it.

References

References

  1. [1]Juan A. Acebrón, Luis L. Bonilla, Conrad J. Pérez Vicente, Felix Ritort, and Renato Spigler. 2005. The Kuramoto model: A simple paradigm for synchronization phenomena. Reviews of Modern Physics 77, 1 (2005), 137–185.DOI
  2. [2]Anthropic. 2024. Building Effective Agents. https://www.anthropic.com/research/building-effective-agents. Published 19 December 2024.
  3. [3]Alberto Bressan. 2000. Hyperbolic Systems of Conservation Laws: The One-Dimensional Cauchy Problem. Oxford Lecture Series in Mathematics and its Applications, Vol. 20. Oxford University Press.
  4. [4]Mert Cemri, Melissa Z. Pan, Shuyi Yang, Lakshya A. Agrawal, et al. 2025. Why Do Multi-Agent LLM Systems Fail? NeurIPS 2025; arXiv:2503.13657.
  5. [5]DeepSeek AI. 2026. DeepSeek Harness: an open-source agent harness where everything is a plugin. https://github.com/deepseek-ai/deepseek-harness.
  6. [6]Yu Deng, Zaher Hani, and Xiao Ma. 2024. Long time derivation of the Boltzmann equation from hard sphere dynamics. arXiv preprint arXiv:2408.07818 (2024).arxiv.org/abs/2408.07818
  7. [7]Martin Gardner. 1970. Mathematical Games: The fantastic combinations of John Conway’s new solitaire game “life”. Scientific American 223, 4 (1970), 120–123.
  8. [8]François Golse and Laure Saint-Raymond. 2004. The Navier–Stokes limit of the Boltzmann equation for bounded collision kernels. Inventiones Mathematicae 155 (2004), 81–161.DOI
  9. [9]Peter K. Haff. 1983. Grain flow as a fluid-mechanical phenomenon. Journal of Fluid Mechanics 134 (1983), 401–430.DOI
  10. [10]David Hilbert. 1902. Mathematical Problems. Bull. Amer. Math. Soc. 8, 10 (1902), 437–479.
  11. [11]Jung and A. Asnacios. 2026. Theory of collective learning in populations of adaptive agents. arXiv preprint arXiv:2607.02171 (2026).arxiv.org/abs/2607.02171
  12. [12]Kai Mei, Xuehan Liu, Shanglee Xue, Qinxxuan Li, Suzhou Ming, Yifei Zhao, et al. 2024. AIOS: LLM Agent Operating System. arXiv:2403.16971arxiv.org/abs/2403.16971
  13. [13]Omarchy Foundation. 2026. Alibaba Cloud Joins as Founding Corporate Patron. https://omarchy.org/news/2026/09/alibaba-cloud-joins-as-founding-corporate-patron/ Announced 22 September 2026; Alibaba Cloud, $3M over three years, and the Qwen Book agent computer with the stated principle “OS as Harness”.
  14. [14]OpenAI. 2026. On the Navier–Stokes Millennium Prize Problem. https://openai.com/index/navier-stokes-solution. Announced 8 September 2026.
  15. [15]Lorenzo Pareschi and Giuseppe Toscani. 2013. Interacting Multigent Systems: Kinetic equations and Monte Carlo methods. Oxford University Press.
  16. [16]Ilya Prigogine. 1978. Time, Structure, and Fluctuations. Science 201, 4358 (1978), 777–785.pubmed.ncbi.nlm.nih.gov/17738519
  17. [17]Craig W. Reynolds. 1987. Flocks, Herds, and Schools: A Distributed Behavioral Model. In Computer Graphics Proceedings of SIGGRAPH ’87, Vol. 21. 25–34.DOI
  18. [18]Yifan Shi, Wei Zhang, and Tianyi Cui. 2026. A Programming Paradigm for Spatiotemporal Composability. https://github.com/cordiverse/paper.arxiv.org/abs/2608.25512
  19. [19]Noah Shinn, Federico Cassano, Ben Labash, Ashwin Gopinath, Karthik Narasimhan, and Shunyu Yao. 2023. Reflexion: Language Agents with Verbal Reinforcement Learning. In Advances in Neural Information Processing Systems 36 (NeurIPS).arxiv.org/abs/2303.11366
  20. [20]Haipeng Wang. 2026. A Kinetic Theory of the Gated Self-Evolving LLM Agent. arXiv:2610.03243 [cs.GR].arxiv.org/abs/2610.03243
  21. [21]Xuening Wu, Shengqin Yin, Yanlan Kang, Xinhang Zhang, Qianya Xu, Zeping Chen, and Wenqiang Zhang. 2025. SGM: A Statistical Model Machine for Risk-Controlled Recursive Self-Modification. arXiv preprint arXiv:2510.1032 (2025).arxiv.org/abs/2510.10232
  22. [22]Yifan Yang, Ziyang Gong, Weiquan Huang, Qihao Yang, Ziwei Zhou, Zisu Huang, Yan Li, Xuemei Gao, Qi Dai, Bei Liu, Kai Qiu, Yuqing Yang, Dongdong Chen, Xue Yang, and Chong Luo. 2026. SkillPot: Executive Strategy for Self-Evolving Agent Skills. arXiv preprint arXiv:2605.23904 (2026).arxiv.org/abs/2605.23904
  23. [23]Yaodong Yang, Rui Luo, Minne Li, Ming Zhou, Weinan Zhang, and Jun Wang. 2018. Mean field multi-agent reinforcement learning. In International Conference on Machine Learning (ICML).arxiv.org/abs/1802.05438
  24. [24]Shunyu Yao, Howard Chen, John Yang, and Karthik Narasimhan. 2022. WebShop: Towards Scalable Real-World Web Interaction with Grounded Language Agents. In Advances in Neural Information Processing Systems.arxiv.org/abs/2207.01206
  25. [25]Xunjian Yin, Xinyi Wang, Liangming Pan, Li Lin, Xiaojun Wan, and William Yang Wang. 2025. Gödel Agent: A Self-Referential Agent Framework for Recursively Self-Improvement. In Proceedings of the 63rd Annual Meeting of the Association for Computational Linguistics (ACL). 27890–27913.

Paper details

vlnc:W_8aeth8ru
Version
v1 · latest
Published
Fresh — added to the Hub within the last 30 days
Added
License
CC BY 4.0
Usage

views · 0

downloads · 0

A view counts once per browser per day, a download each time the PDF is requested.

  • Computer Science

Not recorded for this paper

View statement
Looking for related papers…
Contents