AgentCo-op: Cooperative Multi-Agent Systems

• AgentCo-op is a cooperative multi-agent systems approach that enables retrieval-based synthesis of workflows from reusable skills and explicit coordination interfaces.
• It leverages typed artifact handoffs and local repair mechanisms to ensure robust interoperability among independently developed agents and tools.
• AgentCo-op incorporates coordination primitives from coalition heuristics, decentralized trading, and belief modeling to enhance multi-agent cooperation.

Searching arXiv for papers related to AgentCo-op and interoperable multi-agent workflow design. AgentCo-op denotes a research direction in cooperative multi-agent systems centered on making independently developed agents, tools, and workflows interoperate under explicit coordination mechanisms rather than assuming a monolithic controller. In its most specific usage, AgentCo-op is a retrieval-based synthesis framework that constructs executable multi-agent workflows from reusable skills, tools, external agents, and typed artifact handoffs, then applies bounded self-guided local repair when execution evidence indicates failure (Shen et al., 19 May 2026). More broadly, the literature associated with this theme studies how cooperative behavior can emerge or be engineered through coalition heuristics, trading-based consensus, belief-aware partner modeling, decentralized reinforcement learning, governance layers for shared mutation, and privacy-preserving contribution measurement (Vernon-Bido et al., 2020, Im et al., 5 Feb 2025, Zhang et al., 2023, Matsunaga et al., 29 Jun 2026, Huang, 29 Jun 2026, Xia et al., 24 Dec 2025).

1. Definition and conceptual scope

In the 2026 formulation, AgentCo-op takes as input a task specification

$x = (g, c, r, \Omega),$

where $g$ is the user goal, $c$ the task context, $r$ the operational constraints, and $\Omega$ the task-specific resources, and synthesizes an executable workflow

$$W = Synthesize(x, \mathcal{S}) \triangleq (R, G, \phi, \Pi),$$

with roles $R$, a directed workflow graph $G=(V,E)$, a mapping $\phi$ from roles to attached skills and tools, and an interface protocol $\Pi$ governing messages and typed artifacts (Shen et al., 19 May 2026). The defining move is to replace global workflow search over a scalar benchmark objective with retrieval-based composition of existing components, followed by local repair rather than topology-wide resynthesis.

This specific framework sits naturally beside a wider body of work on cooperative agent organization. In coalition formation, agent-based heuristics have been used to approximate core-stable coalition structures in hedonic games without enumerating the Bell-number space of partitions, showing that local coalition-change rules can often recover members of the core (Vernon-Bido et al., 2020). In noncooperative equilibrium selection, a decentralized trading-based auction can induce consensus on one common equilibrium without direct communication or disclosure of private valuations, again turning individually rational local updates into globally coordinated outcomes (Im et al., 5 Feb 2025). This suggests that AgentCo-op is not only a named framework but also a recurring systems pattern: cooperation is operationalized through explicit interfaces, local admissibility conditions, and structured intermediate representations rather than through unrestricted end-to-end optimization.

A second recurring characteristic is modularity. ProAgent decomposes cooperation into a Knowledge Library, State Grounding, Memory, Planner, Belief Correction, Verificator, and Controller, with zero-shot coordination emerging from intention inference and online belief revision rather than joint training (Zhang et al., 2023). COOPA similarly decomposes operations-research decision support into formulation extraction, confidence evaluation, iterative refinement, solver dispatch, and solver-specific optimizer agents (Li et al., 25 Jun 2026). Across these cases, cooperation is mediated by auditable artifacts—belief traces, typed outputs, source references, or bounded write intents—rather than being left implicit inside a single policy network.

2. Retrieval-based synthesis and typed interoperability

The most explicit AgentCo-op architecture is organized into five stages: Planning, Retrieval, Synthesis, Execution, and Review. It maintains a global library $g$0 containing reference resources, agent skills, tools and APIs, and external agent repositories or existing agent graphs. For a given task, the planner decomposes the goal, inspects available resources, and issues a retrieval plan; synthesis then grounds roles with retrieved components and constructs the workflow graph without running a global topology search (Shen et al., 19 May 2026).

A central mechanism is the typed artifact handoff. Nodes in the workflow do not exchange arbitrary text alone; they produce and consume structured artifacts whose types encode structure and semantics. The Broker validates these types and can transform artifacts when needed. In the serial collaboration case study, TissueAgent produces a MarkerTable, the Broker validates and converts it into the GeneSetInput expected by GeneAgent, and no genes are dropped in the handoff of 53 upregulated markers (Shen et al., 19 May 2026). In the parallel multiome workflow, Seurat and Signac run in separate Docker containers, each yielding typed marker sets that are later intersected or unioned for downstream evaluation (Shen et al., 19 May 2026).

The interoperability problem reappears in systems papers focused on shared mutation control. ATM introduces a specification-grounded governance chain in which task intent, repository scope, write admission, validation, and evidence obligations are bound together; a CID broker decides whether concurrently formed write intents are parallel-safe, require deterministic composition via needs-physical-split, must be serialized, or must take a fail-closed path (Huang, 29 Jun 2026). CoAgent addresses a related systems problem for live shared state by interposing MTPO middleware between agents and tools: reads are order-filtered, writes are applied speculatively in place, and affected readers receive one-way notifications so they can repair only the premises that changed (Lyu et al., 13 Jun 2026). These mechanisms are distinct from AgentCo-op’s typed artifacts, but they serve an analogous purpose: they turn heterogeneous local actions into governed compositions over shared state.

A plausible implication is that typed artifacts and governed mutation are complementary layers of the same interoperability stack. Typed artifacts regulate semantic compatibility between nodes in a synthesized workflow, whereas admission control and speculative undo regulate temporal compatibility when multiple agents touch the same live substrate.

3. Coordination mechanisms across cooperative and noncooperative settings

Several papers linked to AgentCo-op treat coordination as a process problem rather than a static equilibrium concept. In hedonic cooperative games, an agent-based heuristic applies six local routines—join coalitions, exit coalition, create pair coalition, defect coalition, split coalition, and return to individual coalition—to approximate core-stable partitions. On glove games with $g$1, this heuristic ended in a coalition structure in the core in 3363 out of 3500 runs, or 96.1%, while reaching only about 35–36% of distinct core partitions overall (Vernon-Bido et al., 2020). The result indicates that local, greedy deviations can recover some globally stable cooperative outcomes even when exact core computation is NP-complete or co-NP-complete.

TACo addresses an adjacent problem in noncooperative systems with multiple equilibria. It defines a discrete set of choices $g$2, a private cost matrix $g$3, a private valuation vector $g$4, and public offer and pay matrices $g$5 and $g$6, with quasi-linear profit

$g$7

Agents act in round-robin order, choose $g$8, back that choice by updating $g$9 and $c$0, and use cycle detection plus a decrement factor $c$1 on the trading unit $c$2 until the $c$3-termination condition is met (Im et al., 5 Feb 2025). Theoretical analysis proves eventual cycling, bounded profit differences inside cycles, finite termination, and an explicit worst-case step bound. In a waypoint merging game with $c$4 agents and $c$5 candidate equilibria, TACo achieved median optimality gap $c$6, median Gini index about $c$7, median convergence around 53 steps, and maximum 380 steps (Im et al., 5 Feb 2025).

A different cooperative mechanism appears in decentralized local energy markets. There, “implicit cooperation” is formulated as a Dec-POMDP in which agents do not communicate directly but instead observe stigmergic system-level KPIs such as grid balance,

$c$8

supply–demand imbalance, coordination score, liquidity, and congestion. Rewards combine local profit and system-level terms through

$c$9

Under a $r$0 factorial design, APPO-DTDE achieved a coordination score of 91.7% relative to the theoretical centralized benchmark, while DTDE reduced the variance of grid balance by 31% compared to hybrid architectures (Salazar-Pena et al., 17 Feb 2026). This indicates that public aggregate signals can, in some domains, substitute for explicit coordination channels.

Taken together, these works suggest several recurrent coordination primitives within the AgentCo-op orbit: local coalition moves, side-payment-like trading, stigmergic global indicators, and bounded repair. None depends on unrestricted full-information central planning, yet each uses an explicit protocol that constrains how local actions aggregate.

4. Learning, belief modeling, and cooperative adaptation

A substantial branch of the literature emphasizes adaptive partner modeling. ProAgent is designed as a proactive, zero-shot cooperative agent in Overcooked-AI, using a Planner that produces an Analysis and high-level skill, a Belief entry storing inferred teammate intent, a Belief Correction mechanism that either replaces predicted intentions with actual behavior or annotates them as incorrect, and a Verificator that checks skill feasibility and triggers replanning loops (Zhang et al., 2023). In the Cramped Room ablation, full ProAgent scored 204, removing belief dropped performance to 184, and removing both analysis and belief reduced it to 100; removing the Verificator dropped success rate over 100 steps to 20% (Zhang et al., 2023). These results indicate that explicit analysis, intention inference, and feasibility checking materially affect cooperative robustness.

In cooperative MARL, ACPO derives an exact decentralized decomposition of the true joint policy gradient under CTDE by serializing simultaneous actions and introducing beliefs over preceding agents’ actions. The key result is the decomposition

$r$1

which permits per-agent updates whose sum is exactly the joint policy gradient (Matsunaga et al., 29 Jun 2026). Empirically, ACPO outperformed strong baselines on Multi-Robot Warehouse, SMACv2, and MA-MuJoCo, with the gap widening as the number of agents grew (Matsunaga et al., 29 Jun 2026).

Stateful Active Facilitator addresses a different source of difficulty: environments that vary in coordination level and heterogeneity level. It introduces a shared Knowledge Source $r$2 that aggregates agent messages during training and a Policy Pool from which each agent dynamically selects a policy via Gumbel–Softmax attention over learned policy keys (Liu et al., 2022). On HECOGrid, the Knowledge Source improved performance across tasks, especially under higher coordination, while the Policy Pool specifically improved higher-heterogeneity settings (Liu et al., 2022). This separation suggests a useful distinction within AgentCo-op: coordination and heterogeneity may require different architectural responses, even when both manifest as failures of decentralized adaptation.

A related but more domain-specific cooperative design appears in OptAgent for intelligent building operations. Its orchestrator can operate in centralized single-stage, centralized two-stage, or decentralized mode over 11 specialist agents and 72 MCP tools. Across about 3975 runs, centralized two-stage planning achieved tool accuracy 0.69 versus 0.22 for centralized single-stage, agent accuracy 0.72 versus 0.27, and parameter accuracy 0.70 versus 0.34, while also reducing orchestrator token usage relative to centralized single-stage (Jiang et al., 27 Jan 2026). This suggests that, in cooperative tool-use systems, the architecture of coordination can dominate prompt-level improvements.

5. Governance, fairness, and trust in open multi-agent systems

When cooperative systems operate in trustless or high-stakes environments, contribution measurement and mutation control become central. DAO-Agent formalizes a cooperative game over autonomous LLM agents $r$3, with coalition outputs $r$4 and value

$r$5

It uses Shapley values

$r$6

for off-chain contribution measurement, and proves them in zero knowledge against on-chain commitments (Xia et al., 24 Dec 2025). In the crypto-trading case study, on-chain verification gas stayed around $r$7k regardless of coalition size, compared with 367k gas for 4 agents and 28.6M gas for 10 agents under naive on-chain alternatives, yielding up to 99.9% reduction in verification gas costs (Xia et al., 24 Dec 2025). This provides a model for privacy-preserving, auditable reward allocation in decentralized AgentCo-op deployments.

COOPA, although focused on operations research, advances a related concern: source traceability and confidence-backed human verification. Each ParameterDefinition, VariableDefinition, ObjectiveDefinition, and ConstraintDefinition includes a SourceReference, and iterative confidence-based modeling selects the candidate formulation maximizing the minimum confidence across parameters, variables, objective, and constraints (Li et al., 25 Jun 2026). Across three OR benchmarks and eight LLM backbones, COOPA achieved macro-average accuracy 64.8% versus 61.6% for the strongest baseline, with gains up to 6.7 percentage points on GPT-5 (Li et al., 25 Jun 2026). This is not a fairness mechanism in the game-theoretic sense, but it is a governance mechanism in the epistemic sense: every cooperative artifact is tied back to evidence and can be audited element by element.

ATM and CoAgent extend governance to shared code or infrastructure. ATM introduces Task Contracts

$r$8

and three invariants: scope containment, direction stability, and evidence-backed closure (Huang, 29 Jun 2026). CoAgent’s MTPO, by contrast, fixes a serialization preorder at launch, serves each read the order-filtered value, and preserves serializability at quiescence through notifications, self-healing, and undoable tools (Lyu et al., 13 Jun 2026). A plausible implication is that AgentCo-op systems intended for open scientific or software settings require both forms of governance: evidence-backed closure over workflow outputs and concurrency control over the shared substrates those workflows mutate.

6. Applications, empirical results, and limitations

The application space covered by the cited work is unusually broad. In open-world genomics, AgentCo-op synthesized a linear workflow TissueAgent → Broker → GeneAgent → Integrator for MERFISH heart data and a parallel Seurat–Signac workflow for PBMC multiome analysis, with quantitative gains from multimodal marker intersection and union when evaluated against CellMarker 2.0 and PanglaoDB (Shen et al., 19 May 2026). On six standard benchmarks—HotpotQA, DROP, HumanEval, MBPP, GSM8K, and MATH—AgentCo-op with GPT-4o-mini achieved 76.5, 77.2, 90.2, 87.1, 94.4, and 58.2 respectively, yielding the best average score among compared systems under the unified backbone setting while reducing per-task token cost relative to multi-agent baselines such as LLM-Debate and ReConcile (Shen et al., 19 May 2026).

In human-facing cooperation, ProAgent outperformed baselines in all five Overcooked layouts as Player 0 and in three of five as Player 1 when paired with AI agents, and showed an average improvement exceeding 10% over the current state-of-the-art method when partnered with human proxy models (Zhang et al., 2023). In a simpler assistant setting, a DQN-based co-op agent for Space Invaders improved game performance and was perceived as purposeful, though informal testing suggested it often felt more competitive than supportive; the random assistant achieved higher average scores, but the trained assistant was more coherent behaviorally (Krishnan et al., 2021). These results underscore that cooperative quality is not reducible to win rate alone.

In infrastructure and energy domains, OptAgent demonstrated multi-domain, multi-agent coordination over building thermal dynamics, HVAC, DER, controllers, disturbances, simulation, analysis, and comparison, while the decentralized local-energy-market study showed that fully decentralized APPO-DTDE could approach centralized coordination quality and simultaneously deliver a more predictable import-biased load profile (Jiang et al., 27 Jan 2026, Salazar-Pena et al., 17 Feb 2026). In web information access, AgentWebBench showed that decentralized multi-agent coordination generally lags behind centralized retrieval on retrieval-heavy tasks, but the gap shrinks with model scale and can reverse on question answering, where iterative evidence gathering through website-specific content agents can outperform centralized retrieval (Zhong et al., 13 Apr 2026).

The limitations are equally consistent across papers. AgentCo-op’s retrieval-based synthesis does not guarantee global optimality and depends on the quality of resource descriptions, type schemas, and local repair policies (Shen et al., 19 May 2026). Coalition heuristics may become trapped in local maxima and cover only a subset of true core partitions (Vernon-Bido et al., 2020). TACo assumes myopic best-response participation and does not establish full incentive compatibility against farsighted strategic manipulation (Im et al., 5 Feb 2025). ProAgent’s belief model is textual rather than formally probabilistic, and its LLM loops are computationally expensive (Zhang et al., 2023). DAO-Agent still faces $r$9 off-chain Shapley computation, and exact proving cost grows rapidly with coalition size (Xia et al., 24 Dec 2025). CoAgent’s guarantees depend on footprint-declared, undoable tools and on the self-healing assumption for notified agents (Lyu et al., 13 Jun 2026). ATM explicitly does not claim cross-clone governance or broad comparative superiority outside observed single-domain settings (Huang, 29 Jun 2026).

These limitations indicate that AgentCo-op remains less a single settled architecture than a family of coordination techniques. What unifies them is the insistence that cooperation must be engineered through explicit interfaces, bounded local transformations, auditable intermediate artifacts, and domain-appropriate repair mechanisms rather than assumed to emerge automatically from shared model weights or unrestricted global optimization.