Evolutionary Generation of Multi-Agent Systems

• Evolutionary Generation of MAS (EvoMAS) is a framework that models multi-agent system design as an evolutionary search over structured configurations.
• It employs techniques such as mutation, crossover, and selection on graph-based, hierarchical, and controller representations to boost adaptability.
• EvoMAS supports online, test-time, and continuous self-evolution, improving complex tasks like reasoning, tool use, and agent collaboration.

Searching arXiv for the cited EvoMAS papers to ground the article in current preprints. Searching arXiv for "Evolutionary Generation of Multi-Agent Systems". Evolutionary Generation of Multi-Agent Systems (EvoMAS) is a family of approaches that treats the construction, adaptation, and optimization of multi-agent systems as an evolutionary search problem over structured system designs rather than as one-shot manual engineering. In this view, the evolving object may be a population of agents, a controller program, a hierarchical organization, a communication graph, a declarative configuration, or a workflow scaffold. Recent LLM-based formulations make this perspective explicit by searching over prompts, models, tools, roles, and topologies, while earlier work established closely related ideas for evolving organizational structures, online controller adaptation, eco-evolutionary populations, and stability analysis of evolving agent populations (Hu et al., 6 Feb 2026, He et al., 22 Apr 2026, Shen et al., 2014, Nunes et al., 2014, Hamon et al., 2023, Wilde et al., 2011).

1. Historical lineage and problem formulation

EvoMAS predates contemporary LLM-based agent systems. Earlier work studied evolving agent populations as stochastic systems whose number of agents can vary according to fitness, extending Markov-chain stability analysis from fixed-population MAS to evolving populations and introducing an entropy-based degree of instability, $\delta = -\sum p_i^\infty \log_N(p_i^\infty)$ (Wilde et al., 2011). In parallel, hierarchical MAS organization was cast as a genetic search problem over tree-shaped structures, motivated by the observation that a large-scale system may have “billions of possible ways of organization,” making exhaustive search impractical (Shen et al., 2014). Online controller evolution was also explored through Cartesian Genetic Programming and a $(1+\lambda)$ evolutionary strategy with $\lambda = 4$, embedding breeding and program selection directly into runtime interaction rather than separating evolution from execution (Nunes et al., 2014). A different biological line of work demonstrated continuous neuroevolution in a large non-episodic multi-agent environment where neither the environment nor the population is reset, and where reproduction and death depend on internal physiology rather than an external episodic reward loop (Hamon et al., 2023).

Contemporary LLM-based EvoMAS inherits this broad evolutionary view but shifts the search target from low-level controllers to high-level agentic system design. The central motivation is consistent across papers: effective MAS performance depends on many interacting components, including agent roles, prompts, tools, memory, communication rules, workflow structure, and organizational topology, which together define a high-dimensional, discrete, and difficult-to-engineer search space (He et al., 22 Apr 2026, Hu et al., 6 Feb 2026). The 2026 paper explicitly titled “Evolutionary Generation of Multi-Agent Systems” formulates MAS generation as structured configuration generation in configuration space, positioning EvoMAS between brittle code-generation approaches and rigid template-based methods (Hu et al., 6 Feb 2026). Related LLM-era frameworks then specialize this general idea in different ways: automatic extension of a specialized agent into a collaborating population via mutation, crossover, and selection (Yuan et al., 2024); closed-loop self-improvement through textual graph edits (He et al., 22 Apr 2026); test-time co-evolution of topology and capability during a single query (Xu et al., 10 May 2026); and collaborative self-evolution from distributed execution traces (Hao et al., 28 May 2026).

A persistent theme across this history is that EvoMAS is not limited to offline synthesis. Some methods evolve organizational structure before deployment, some evolve communication or capability during inference, and some maintain continuous adaptation throughout ongoing interaction. This suggests that EvoMAS is best understood as a unifying systems perspective rather than a single algorithmic template.

2. Search spaces and system representations

The defining methodological choice in EvoMAS is how the multi-agent system is represented as an evolvable object. In the configuration-space formulation, a MAS is represented as

$C = (G, \{A_i\}_{i=1}^k, V_{in}, V_{out}),$

where $G = (V, E)$ is an acyclic communication graph, each agent $A_i = (b_i, p_i, T_i)$ contains a backbone model, a system prompt template, and a tool set, and $V_{in}, V_{out}$ specify distinguished input and output agents (Hu et al., 6 Feb 2026). Because this representation is declarative, the generator searches over structured specifications rather than over raw executable code.

Other recent frameworks use graph-native abstractions. TPGO introduces the Textual Parameter Graph,

$\mathcal{G} = (\mathcal{V}, \mathcal{E}),$

whose nodes are Role nodes, Logic nodes, and Tool nodes, and whose directed edges capture role-to-logic, logic-to-tool, and cross-agent dependencies (He et al., 22 Apr 2026). TacoMAS formalizes inference itself as online graph adaptation, with

$$\mathcal G_t = (T_t,\Phi_t), \qquad T_t = (\mathcal V_t,\mathcal E_t),$$

so that the node set, edge set, and capability states all become time-varying quantities during execution (Xu et al., 10 May 2026). ST-EVO pushes this further to a spatio-temporal graph,

$$\mathcal{G} = \{\mathcal{G}_t\}_{t=1}^T = \{(\mathcal{V}, \mathcal{E}_t, \mathrm{A}_t)\}_{t=1}^T,$$

in which topology evolves dialogue by dialogue rather than remaining query-static (Wu et al., 16 Feb 2026). AgentNet also adopts a graph representation, $(1+\lambda)$0, but emphasizes a decentralized DAG whose edges are strengthened or pruned by local collaboration success (Yang et al., 1 Apr 2025).

Earlier EvoMAS representations were equally structural, though at different abstraction levels. The hierarchical genetic algorithm encodes a hierarchical organization with $(1+\lambda)$1 leaf nodes and maximum depth $(1+\lambda)$2 as an integer array $(1+\lambda)$3, where each $(1+\lambda)$4 gives the level at which adjacent leaves start to separate (Shen et al., 2014). Cartesian Genetic Programming represents controllers as feed-forward acyclic directed graphs encoded by real-valued genes $(1+\lambda)$5 (Nunes et al., 2014). Eco-evolutionary neuroevolution instead evolves recurrent neural controllers with local perception and an internal energy state, embedding the agent within a persistent ecological process rather than representing only an abstract task workflow (Hamon et al., 2023).

These representations differ in semantics, but they share a core property: each makes a multi-agent system legible to variation and selection. In recent LLM-based work, the shift from monolithic prompts to structured graphs and configurations is especially consequential because it permits local edits, architectural recombination, and topology-aware diagnosis rather than only surface-level prompt rewriting.

3. Evolutionary operators, feedback signals, and memory

Classical EvoMAS uses familiar evolutionary operators. EvoAgent starts from a pre-defined specialized agent framework and iterates through Initialization, Crossover + Mutation, Selection, and Results update, using an LLM as the “evolution engine” and an LLM-based quality check to decide whether a candidate expert should be retained or discarded (Yuan et al., 2024). The configuration-space EvoMAS framework also uses Select, Mutate, Crossover, and Consolidate, but constrains mutation to change exactly one component type per step—prompts, model IDs, tools, or topology—and constrains crossover so that the entire topology is inherited from one parent while agent-level attributes are recombined from both (Hu et al., 6 Feb 2026). The hierarchical genetic algorithm replaces generic GA operators with hierarchical crossover plus repair and mutation of small perturbation, so that exchanged or perturbed genome segments correspond to meaningful tree substructures rather than arbitrary slices (Shen et al., 2014).

Recent LLM-centered systems broaden the notion of evolutionary feedback. TPGO defines “textual gradients” as structured natural-language feedback extracted from execution trajectories. Negative and positive gradients are generated from reasoning traces, tool calls, and inter-agent communication, then negative gradients are embedded and clustered with DBSCAN so that updates target recurring failure modes such as tool misuse or weak verification rather than one-off mistakes (He et al., 22 Apr 2026). ST-EVO adds an uncertainty-aware reward based on predictive entropy and VarEntropy, using

$(1+\lambda)$6

and couples this with retrieval of successful past scheduling trajectories from memory (Wu et al., 16 Feb 2026). AgentNet updates connection weights after each task by

$(1+\lambda)$7

then prunes edges below a threshold, so successful collaboration literally changes the network (Yang et al., 1 Apr 2025). Meta-Team treats distributed execution experience itself as the substrate of evolution, preserving each agent’s local trace, cross-agent messages, and shared artifacts, then applying updates at agent, interaction, and team levels rather than collapsing everything into a single centralized analyzer (Hao et al., 28 May 2026).

Memory increasingly acts as a meta-evolutionary mechanism. GRAO in TPGO stores problem context $(1+\lambda)$8, proposal $(1+\lambda)$9, and effectiveness $\lambda = 4$0, retrieves semantically similar historical cases, and ranks them by effectiveness to condition future edits (He et al., 22 Apr 2026). The configuration-based EvoMAS framework consolidates each task’s evolution trace into an experience record $\lambda = 4$1 for future retrieval by task similarity (Hu et al., 6 Feb 2026). AgentNet stores only local trajectory fragments in router and executor memories, emphasizing privacy-preserving and compact specialization (Yang et al., 1 Apr 2025). This suggests that contemporary EvoMAS increasingly combines evolutionary variation with case-based transfer, so that search is conditioned not only by current reward but also by accumulated design experience.

4. Online, test-time, and distributed self-evolution

A major distinction within EvoMAS is whether evolution occurs before deployment, during test-time inference, or continuously during ongoing interaction. TPGO presents a closed-loop, self-improving optimization process in which a parser decomposes prompts into a Textual Parameter Graph, the MAS is executed, textual gradients are generated, recurring failures are clustered, graph edits are proposed and validated, and the resulting experience is stored for future optimization (He et al., 22 Apr 2026). The “evolutionary” aspect is explicit: repeated cycles of selection pressure through validation, variation through graph edits, and memory-based adaptation through GRAO.

TacoMAS makes evolution an intra-query phenomenon. Its central claim is that effective test-time evolution requires jointly adapting capability and topology, but on different time scales:

$\lambda = 4$2

where capability updates occur every execution round while topology changes only every $\lambda = 4$3 rounds (Xu et al., 10 May 2026). The fast loop scores each agent’s contribution and writes prompt or memory refinements back into its capability state, while the slow loop performs bounded birth-death and edge-edit operations. The paper’s interpretation of this fast-slow design as a replicator-mutator process moving toward a task-conditioned stable equilibrium gives EvoMAS a clear game-theoretic and dynamical-systems reading.

Distributed self-evolution is emphasized by AgentNet and Meta-Team. AgentNet rejects a central orchestrator and lets each agent decide locally whether to forward, split, or execute a task, while topology evolves through local success-weighted edge updates and capability vectors are updated over time (Yang et al., 1 Apr 2025). Meta-Team argues that a MAS should not only execute as a team, but also evolve as a team. After each task, each agent reflects on its own local execution chain, exchanges evidence with teammates, and contributes to L1 agent-level, L2 interaction-level, and L3 team-level evolution (Hao et al., 28 May 2026). The contrast with centralized analyzers is substantive rather than stylistic: the paper reports that collaborative experience organization outperforms both global flattening and purely local analysis on failure attribution.

Continuous online evolution also remains present outside LLM agents. The CGP-based online generation algorithm maintains a current behavior population, executes selected programs for a window of $\lambda = 4$4 steps, evaluates energy samples, breeds a new generation when the selector is done, and reloads the population (Nunes et al., 2014). Non-episodic neuroevolution demonstrates a still stronger form of continuity: reproduction, death, resource consumption, and resource regrowth all occur in a single uninterrupted ecological loop with no environment or population reset (Hamon et al., 2023). In this broader perspective, LLM-based test-time evolution is one branch of a larger EvoMAS agenda concerned with systems that keep changing while they act.

5. Empirical evidence and application domains

Recent empirical work shows that EvoMAS methods operate across reasoning, tool use, software engineering, virtual world generation, and ecologically grounded simulation. Selected reported results from TPGO, TacoMAS, EvoMAS, Meta-Team, EvoAgentX, and EvoAgent indicate that gains come from changing coordination structure, prompts, tools, and organizational design rather than merely substituting a stronger base model (He et al., 22 Apr 2026, Xu et al., 10 May 2026, Hu et al., 6 Feb 2026, Hao et al., 28 May 2026, Wang et al., 4 Jul 2025, Yuan et al., 2024).

Framework	Benchmarks	Reported result
TPGO	MCP-Universe, GAIA	ReAct (GPT-4.1) 30.96% $\lambda = 4$5 38.82%; ReAct (DeepSeek-V3.2) 34.77% $\lambda = 4$6 41.86%; MiroFlow (GPT-5) 73.8% $\lambda = 4$7 81.6%; average task time 4014s $\lambda = 4$8 1765s
TacoMAS	finance, browsecomp, plancraft, workbench	0.767, 0.745, 0.887, 0.824; average improvement of 13.3% over the strongest baseline
EvoMAS	BBEH, WorkBench, SWE-Bench-Verified	58.7% on BBEH, 48.9% on WorkBench, and 79.1% on SWE-Bench-Verified with Claude-4.5-Sonnet
Meta-Team	six long-horizon benchmarks	62.7 overall average in the main table, versus 56.1 for the hand-crafted MAS and 56.4 for MASFly
EvoAgentX	HotPotQA, MBPP, MATH, GAIA	+7.44% HotPotQA F1, +10.00% MBPP pass@1, +10.00% MATH solve accuracy, and overall accuracy improvement of up to 20.00% on GAIA
EvoAgent	Logic, ScienceWorld, TravelPlanner	GPT-4 logic 77.00 vs 65.50 CoT; ScienceWorld GPT-4 27.97 $\lambda = 4$9 30.42; TravelPlanner GPT-4 final pass rate 7.2 with EvoAgent $C = (G, \{A_i\}_{i=1}^k, V_{in}, V_{out}),$0 versus 2.2 for Direct/CoT

The application distribution is correspondingly broad. EvoAgent emphasizes automatic extension of a specialized agent into a task-specialized expert population and reports gains on Logic Grid Puzzle, Trivia Creative Writing, Codenames Collaborative, MMMU, ScienceWorld, and TravelPlanner (Yuan et al., 2024). EvoAgentX presents a unified open-source platform in which TextGrad and MIPRO refine prompts and AFlow restructures workflows, supporting reasoning, code generation, math, and real-world agent tasks (Wang et al., 4 Jul 2025). The configuration-space EvoMAS paper targets BBEH, SWE-Bench, and WorkBench, and emphasizes not only accuracy but also very high execution rates, including 98.9% on BBEH and 98.4% on SWE-Bench-Verified (Hu et al., 6 Feb 2026). Outside LLM benchmarks, eco-evolutionary neuroevolution reports sustainable collective foraging, population-resource oscillations, and two distinct behavioral types—opportunistic travelers and low-movement sustainable foragers—in a large persistent grid world (Hamon et al., 2023). A human-agent interactive evolutionary system for procedural city generation reports that a Human-Based Genetic Algorithm required 52 total generations but only 18 interactive generations on average, reducing average time from 18 minutes in IGA to 12 minutes while targeting designer-preferred virtual city layouts (Kruse et al., 2016).

A common interpretation of these results is that performance gains arise when evolution operates on the correct structural unit. TPGO’s ablations show that removing the Textual Parameter Graph, removing structural graph edits, or replacing semantic clustering with random clustering hurts performance (He et al., 22 Apr 2026). TacoMAS reports that if topology changes too frequently trajectories become unstable, whereas if topology is frozen specialization stalls (Xu et al., 10 May 2026). EvoMAS reports that evolved systems often become smaller and more specialized than their ancestors, with agent counts dropping by about 70% and edge density by about 95% in some cases (Hu et al., 6 Feb 2026). These patterns imply that successful evolution is often compressive and structural rather than simply additive.

6. Theory, stability, misconceptions, and open problems

Theoretical work on EvoMAS focuses on convergence, stability, and controllability. The Markov-chain framework for evolving MAS defines stability as existence of a non-uniform limiting distribution over states or macro-states and introduces entropy of the limit probabilities as a macroscopic degree of instability (Wilde et al., 2011). In simulation, the system reached the maximum macro-state $C = (G, \{A_i\}_{i=1}^k, V_{in}, V_{out}),$1 by about generation 178, remained there after about generation 482, and yielded $C = (G, \{A_i\}_{i=1}^k, V_{in}, V_{out}),$2 at $C = (G, \{A_i\}_{i=1}^k, V_{in}, V_{out}),$3; mutation rate, not crossover rate, was reported as the main driver of instability (Wilde et al., 2011). TacoMAS supplies a more modern dynamical result, interpreting fast capability refinement and slow topology mutation as a replicator-mutator process contracting toward a neighborhood of a task-conditioned stable equilibrium under bounded-noise and biased-edit assumptions (Xu et al., 10 May 2026). ST-EVO contributes a different form of stability control through uncertainty-aware scheduling rewards and reports that stability improves by about 40% with the uncertainty-aware reward (Wu et al., 16 Feb 2026).

Several misconceptions are corrected by the literature. EvoMAS is not synonymous with flat prompt tuning: TPGO explicitly contrasts node content refinement and structural graph modification with prompt-only optimization, and its case study inserts a “Constraint Validation Protocol” node rather than merely rewriting a monolithic prompt (He et al., 22 Apr 2026). EvoMAS is also not equivalent to raw code generation: the configuration-space formulation was proposed precisely because code-generation approaches can suffer executability failures, syntax/runtime brittleness, and fragile behavior, while rigid templates sacrifice expressiveness (Hu et al., 6 Feb 2026). Nor does EvoMAS imply that more agents or more computation automatically improve performance. EvoAgent reports that larger populations do not always help and that gains can plateau beyond a certain number of iterations (Yuan et al., 2024). AgentNet reports improved performance with more agents and executor pool size but also diminishing returns (Yang et al., 1 Apr 2025). The CGP online-evolution study reports that continuous evolution completely ruined genotype diversity compared with the initial population and did not demonstrate long-term adaptation in that setup (Nunes et al., 2014).

Current limitations are correspondingly heterogeneous. Recent LLM-based frameworks remain dependent on backbone-model quality, long-context handling, and expensive evaluation loops; EvoAgentX states that memory optimization is under active development, while Meta-Team explicitly does not update base LLM parameters and evolves only scaffolds and organization rather than tool APIs or infrastructure (Wang et al., 4 Jul 2025, Hao et al., 28 May 2026). AgentNet identifies routing scalability in very large heterogeneous systems as an open problem (Yang et al., 1 Apr 2025). Eco-evolutionary neuroevolution notes that movement does not incur energy costs and that richer ecological dynamics remain to be added (Hamon et al., 2023). These constraints indicate that EvoMAS remains an active research program rather than a settled engineering recipe.

Taken together, the literature supports a broad but precise characterization of EvoMAS: it is the study of how multi-agent systems can be generated and improved by iterative variation, feedback, selection, and memory over structured system representations. In recent work, the most consequential shift is from evolving isolated agents to evolving coordination itself—roles, communication pathways, workflows, constitutions, and organizational structure—often during execution rather than only before it.