Group-Evolving Agents: Dynamics and Applications

Updated 25 February 2026

Group-evolving agents are multi-agent systems that evolve collectively through interaction and experience sharing, enabling open-ended self-improvement.
They employ evolutionary operators such as crossover, mutation, and graph-level workflow optimization to enhance emergent cooperation and performance.
Applications include automated scientific discovery, workflow synthesis, and social simulation, where dynamic group adaptation outperforms isolated agent evolution.

Group-evolving agents are multi-agent systems in which evolution—structural, behavioral, or cognitive—occurs at the collective (group, workflow, or organizational) level. These systems exploit intra-group experience sharing and interaction to achieve open-ended self-improvement, enhanced problem-solving, and emergent complexity beyond what isolated agents or tree-structured evolution can provide. This paradigm contrasts with traditional evolutionary methods operating on isolated agents or rigidly branched phylogenies, by allowing the explicit exchange and reuse of experience within the group throughout evolution. Group-evolving agents are foundational in domains such as automated scientific discovery, multi-agent reinforcement learning, social simulation, and autonomous workflow generation.

1. Formal Models and Core Architectures

Group-evolving agent systems represent evolutionary units as explicit agent groups or networks, each maintaining internal dependencies and possibly experiencing structural adaptation across time. The formalization typically involves one or more of the following:

Directed graphs or trees: Agents (nodes) with defined inter-agent edges, as in the tree-structured group architectures of S-Agents (Chen et al., 2024) and workflow graphs in EvoAgentX (Wang et al., 4 Jul 2025). Each agent $a_i$ is parameterized by prompt, memory, and action modules, with group $\mathcal{A} = \{a_1,\dots,a_n\}$ communicating via a graph $\mathcal{G}=(V,E)$ .
Hierarchical multiway aggregation: Used in large-scale social simulations, "group agents" $G_g$ represent aggregates with a layered structure to enable scalable simulation while capturing population diversity (Zhang et al., 4 Jun 2025).
Markov and replicator dynamics: Group-level population state vectors $\xi^t=(\xi_1^t,\dots,\xi_n^t)$ or mixed strategies $x_i \in\Delta^{|A_i|-1}$ evolve according to population-level stochastic or replicator processes, especially in social or evolutionary game-theoretic contexts (Skoulakis et al., 2020, Wilde et al., 2011).
Co-evolutionary preference graphs: In co-evolving and failure-driven agent frameworks, multiple agents jointly optimize by exchanging trajectories, typically via structured pairwise or groupwise preference optimizations (Jung et al., 27 Nov 2025).

In each case, group identity is not static but is maintained through dynamic communication, experience exchange, and possibly explicit split/merge operations, enabling the population to track and exploit emergent diversity.

2. Evolutionary Algorithms and Open-Ended Improvement

Evolution in group-evolving agent frameworks is enacted through a sequence of operators that act either on agent-role/skill configurations, workflow structures, or internal policy weights. Canonical evolutionary mechanisms include:

Crossover and mutation on agent “genomes”: In EvoAgent (Yuan et al., 2024), an agent’s configuration (prompt, role, skill) is treated as a discrete genomic entity, and LLM-based evolutionary operators (Evo_Crossover, Evo_Mutation) are used to generate candidate agents. Selection is mediated via LLM quality checks, and parent–offspring transitions are recorded across evolutionary generations.
Graph-level workflow optimization: EvoAgentX (Wang et al., 4 Jul 2025) generalizes this with group-level optimization over workflows $(\mathcal{G}, P, \Theta)$ , using optimizers like TextGrad (gradient-based prompt optimization), AFlow (workflow graph mutation–crossover), and MIPRO (demonstration search). The entire workflow, including its agent structure, evolves as a single entity.
Co-evolutionary learning from hard negatives: Co-Evolving Agents (Jung et al., 27 Nov 2025) involve target and failure agents jointly training; group-level extension involves $M\times M$ pairwise trajectory exchanges, enhancing diversity and modular specialization but introducing challenges in stability and role allocation.
Intra-group experience sharing: The group-evolving paradigm (GEA) explicitly leverages shared experience within the evolving unit, greatly increasing utilization of exploratory diversity and sustaining progress beyond conventional isolated-branch methods (Weng et al., 4 Feb 2026).
Group adaptation via feedback and replanning: S-Agents dynamically adapt workloads across their organizational tree using progress monitoring and immediate reassignment rather than via externally imposed evolutionary iterations (Chen et al., 2024).

These methods are typically instantiated alongside structured evaluation (fitness functions), agent selection, and global or group-level performance maximization, subject to real-world constraints.

Group-evolving agent dynamics emerge when population structures or social networks are subject to evolutionary pressures, continuous adaptation, or learning-based updating:

Evolutionary population MARL: In large-scale MARL, agents in evolutionary-scale populations (e.g., $N=200,000$ ) evolve strategies via parallelized policy gradient or opponent-learning-aware updates, showing phase transitions in outcome structure (e.g., cooperation in Stag–Hunt only above critical LOLA presence) (Bouteiller et al., 2024).
Co-evolving agents and adaptive games: Double evolutionary processes, where both agents and the games (environments) co-evolve, lead to recurrent dynamics, information-theoretic invariants, and long-run convergence to Nash equilibria of average games, as shown via replicator equations in endogenously evolving zero-sum polymatrix games (Skoulakis et al., 2020).
Stability in evolving populations: Markov models of evolving multi-agent systems rigorously define and analyze stability (stationarity and occupancy probabilities), with entropy-based metrics quantifying instability and transition regimes under different evolutionary parameters (Wilde et al., 2011).
Social boundary and stance formation: Human–agent hybrid societies exhibit endogenously evolving stances and community boundaries, with group-level differentiation, boundary-strength quantification (e.g., modularity $Q$ , silhouette score), and dynamic adaptation to discourse strategies or interventions (Zhang et al., 24 Aug 2025).
Group agents in large-scale networks: Hierarchically organized "group agents" aggregate individual behavior, propagating state, perception, and memory over events, enabling tractable simulation of billions-scale social systems while tracking intergroup variation (Zhang et al., 4 Jun 2025).

These dynamics illustrate the broader principle: group-level learning rules and structural adaptation mechanisms substantially alter the evolutionary and behavioral trajectory of the agentic population, enabling properties such as emergent cooperation, rapid convergence, sustained diversity, or stable self-organization as required by the underlying environment.

Group-evolving agent paradigms underpin a range of advanced multi-agent applications:

Automated scientific discovery: ASCollab (Liu et al., 8 Oct 2025) implements LLM-based research agents that self-organize into evolving peer-review and collaboration networks, continually accumulating findings on the diversity–quality–novelty frontier. Agents with heterogenous exploration/exploitation tendencies dynamically rewire their collaboration graphs, optimizing both novelty and acceptance rates.
Evolutionary workflow synthesis: EvoAgentX (Wang et al., 4 Jul 2025) automatically generates and optimizes multi-agent workflows for reasoning, code synthesis, and real-world planning tasks. Evolution acts directly on agent roles, tool configurations, workflow graphs, and memory policies, delivering large empirical performance gains over static or singly-optimized frameworks.
Open-ended task solving in open worlds: S-Agents (Chen et al., 2024) demonstrate robust, tree-structured, asynchronously collaborating groups for open-ended, multi-agent tasks such as collaborative building or resource-collection in Minecraft, exploiting dynamic coordination without fixed workflows.
Social network and crowd simulation: GA-S $^3$ (Zhang et al., 4 Jun 2025) employs group agents to efficiently simulate macro-level traffic and sentiment response to online events, providing a benchmark for realistic, large-scale social behavior modeling.

Performance metrics in these systems range from group-level task accuracy (e.g., HotPotQA F1), efficiency (time-to-completion, prompt call counts), social metrics (intergroup trust, modularity), to emergent population statistics (meta-score distributions, diversity indices).

A core innovation in group-evolving frameworks is the explicit treatment of intra-group experience and diversity as evolutionary resources:

Utilization of exploratory diversity: Group-evolution (e.g., GEA (Weng et al., 4 Feb 2026)) is empirically shown to convert early-stage behavioral diversity into sustained, long-term performance advantages.
Experience sharing: Groups allow explicit reuse of experience (including failures) to inform future strategy (see hard-negative sharing in (Jung et al., 27 Nov 2025)), accelerating improvement and robustness.
Transferability and bug-fix efficiency: Evolution at the group level yields improved transfer across agent architectures and reductions in mean iterations to framework-level bug-fix compared to isolating self-evolving agents (e.g., 1.4 iterations vs 5 on coding benchmarks) (Weng et al., 4 Feb 2026).
Maintenance of diversity: Even when group evolution ensures convergence to high-performing solutions, population-level structural and behavioral diversity can be tuned through the incorporation of heterogeneous agent policies, preference graphs, or differential selection pressures (see population-level dynamics in (Bouteiller et al., 2024, Liu et al., 8 Oct 2025)).
Limitations and open questions: With increasing group size, stability and convergence properties may become problematic—openly acknowledged in the literature as open research areas regarding regularization, credit allocation, and interaction protocol design (Jung et al., 27 Nov 2025, Wang et al., 4 Jul 2025).

6. Theoretical Guarantees, Stability, and Network Generalizations

Rigorous theoretical analysis in group-evolving agent systems addresses stability, convergence, and complexity:

Equilibrium guarantees: For rescaled zero-sum polymatrix games, the time-average agent behavior and utility converge to the Nash equilibrium of the time-average game, computable in polynomial time (Skoulakis et al., 2020).
Stability via entropy metrics: Stability in evolving populations is assessed via the limiting occupancy probabilities and their entropy (degree of instability), enabling diagnosis of concentration vs. dispersal in macro-states (Wilde et al., 2011).
Co-evolution recurrence and conservation laws: Continuous-time replicator dynamically preserves information-theoretic invariants and exhibits Poincaré recurrence, ensuring neither extinction nor trivial uniformity dominates the evolutionary trajectory (Skoulakis et al., 2020).
Complexity and scalability: Group-evolving optimization is polynomial (or quasi-polynomial) in workflow size and population, with practical scaling to workflows of 20+ nodes and experimental runs leveraging hardware parallelism (Wang et al., 4 Jul 2025, Bouteiller et al., 2024). Stability and convergence, however, may deteriorate with unregulated group expansion or insufficient diversity maintenance.

These properties ensure that group-evolving frameworks combine empirical efficacy with theoretical soundness, albeit with system-specific stability, convergence, and diversity tradeoffs that require careful architectural and protocol design.

7. Experimental Validation and Benchmarks

Empirical validation of group-evolving agents utilizes benchmarks encompassing reasoning, programming, planning, scientific discovery, and social network prediction:

System	Benchmark/task	Group-evolution improvement	Reference
GEA	SWE-bench Verified, Polyglot	71.0% vs. 56.7%; 88.3% vs. 68.3%	(Weng et al., 4 Feb 2026)
EvoAgentX	HotPotQA F1, MBPP pass@1, MATH	+7.44% F1; +10.00% pass@1; +10.00% solve accuracy	(Wang et al., 4 Jul 2025)
S-Agents	Minecraft collaborative tasks	Parallelization, fallback, leader–worker division	(Chen et al., 2024)
ASCollab	Cancer cohort discovery (TCGA)	Q=4.1, N=4.0 for top-25 findings, sustained D=0.75	(Liu et al., 8 Oct 2025)
GA-S $^3$	Online event SNB (social simulation)	MAPE=16.48% (vs. ~70% baseline), Z=0.81	(Zhang et al., 4 Jun 2025)

This multi-domain experimental evidence underpins the core claim: group-evolving agent paradigms deliver robust, transferable, and scalable improvements relative to both static frameworks and classical agent-level evolution, provided that intra-group diversity management and experience sharing are incorporated into the evolutionary protocol.

References:

(Chen et al., 2024) S-Agents: Self-organizing Agents in Open-ended Environments (Bouteiller et al., 2024) Evolution of Societies via Reinforcement Learning (Zhang et al., 4 Jun 2025) GA-S $^3$ : Comprehensive Social Network Simulation with Group Agents (Wang et al., 4 Jul 2025) EvoAgentX: An Automated Framework for Evolving Agentic Workflows (Liu et al., 8 Oct 2025) Hypothesis Hunting with Evolving Networks of Autonomous Scientific Agents (Jung et al., 27 Nov 2025) Co-Evolving Agents: Learning from Failures as Hard Negatives (Weng et al., 4 Feb 2026) Group-Evolving Agents: Open-Ended Self-Improvement via Experience Sharing (Skoulakis et al., 2020) Evolutionary Game Theory Squared: Evolving Agents in Endogenously Evolving Zero-Sum Games (Wilde et al., 2011) Stability of Evolving Multi-Agent Systems (Yuan et al., 2024) EvoAgent: Towards Automatic Multi-Agent Generation via Evolutionary Algorithms (Zhang et al., 24 Aug 2025) Evolving Collective Cognition in Human-Agent Hybrid Societies: How Agents Form Stances and Boundaries (Gajamannage et al., 2015) Identifying manifolds underlying group motion in Vicsek agents