Source-linked AI summary
Co-Evolution in Agentic Systems: Toward Self-Directed Evolution Beyond Human Design
Qing Zong, Jiayu Liu, Junhao Shen, Zecong Tang, Linsi Wu, Yuxuan Liu, Rui Wang, Zhaowei Wang, Weiqi Wang, Cheng Qian, Xiusi Chen, Yangqiu Song
TL;DR
Single-agent self-evolution can remain bounded by fixed tasks and feedback, while co-evolution offers a broader path for adaptive improvement. This survey organizes the field with a three-stage taxonomy spanning agent–agent, agent–environment, and meta co-evolution, and identifies evaluation, scaling, and governance challenges.
Problem
Single-entity self-evolution can remain bounded by fixed external conditions, while co-evolution in agentic systems lacks a focused survey.
Method
The survey defines co-evolution and organizes its literature with a progressive taxonomy spanning agents, environments, and the evolution mechanism.
Results
The taxonomy traces expanding evolutionary freedom from agent–agent adaptation to agent–environment adaptation and ultimately an evolvable evolution mechanism.
Takeaways & Limitations
Future progress lies in agents that continually improve through co-evolution, supported by evaluation and governance that keep the process reliable and controllable.
Takeaways & Limitations
Meta co-evolution remains at an early stage, with limited work meeting the survey’s definition and much discussion relying on uncoupled single-entity precursors.
Abstract
from arXiv · showhide
Agentic systems are increasingly expected to improve after deployment, yet single-entity self-evolution is often bounded by a static learning context, such as fixed tasks and feedback. This survey focuses on co-evolution in agentic systems, a multi-component form of self-evolution in which multiple agents and their environment impose adaptive pressure on one another. To organize existing papers, we propose a progressive three-stage taxonomy that traces how the system gradually sheds human-engineered constraints. Agent--Agent Co-Evolution studies how agents adapt through dynamic peers, including adversarial, collaborative, and organizational adaptation. Agent--Environment Co-Evolution extends this loop to adaptive tasks, feedback, and interaction spaces that change with the agents. Meta Co-Evolution further explores the possibility of making the evolution mechanism itself evolvable. We also discuss open challenges in evaluating such systems, scaling them across multiple components, and keeping increasingly autonomous evolutionary processes safe and controllable. This survey provides a unified foundation for building robust and open-ended agentic systems that can improve beyond fixed human-designed paths.
1 Introduction
The introduction frames co-evolution as a multi-component extension of self-evolution that addresses the limits of improvement under fixed external conditions. It presents a progressive three-stage taxonomy—from agent adaptation, to environmental adaptation, to evolvable evolution mechanisms—and identifies evaluation, scaling, and governance as open challenges.
- Motivation: Single-entity self-evolution updates one agent from experience, feedback, and failures, but can remain bounded by fixed external conditions.Examples include refining the model backbone, updating memory, and improving skills.
- Motivation: Co-evolution instead makes counterparts adaptive, enabling agents to evolve with peers, environments, or the evolution mechanism itself.Adaptive counterparts include opponents, collaborators, tasks, feedback, and interaction spaces.
- Contributions: The survey offers a focused account of co-evolutionary agentic systems and distinguishes mutual adaptation from mere interaction or information exchange.It organizes prior literature around co-evolution rather than treating it only as a subtheme of self-evolving agents.
- Taxonomy: The survey proposes a progressive three-stage taxonomy that expands evolutionary freedom from agents alone, to agents and environments, and ultimately to an evolvable evolution mechanism.The stages are Agent–Agent Co-Evolution, Agent–Environment Co-Evolution, and Meta Co-Evolution.
- Challenges: It also highlights open challenges in evaluating, scaling, and governing increasingly autonomous co-evolutionary systems.These challenges concern systems whose evolutionary freedom expands beyond fixed human-engineered constraints.
2 Preliminaries
The preliminaries model an agentic system as an evolving agent collective and environment, with co-evolution defined as coupled adaptation among evolving units. They organize this process into three stages that progressively remove human-engineered constraints, culminating in evolution of the evolution mechanism itself.
- System Representation: An agentic system S = (A, E) comprises an agent collective A and environment E, while each agent ai = (mi, hi) combines a model backbone with an optional harness.The harness may include memory, tools, skills, prompts, or workflows, or none of these components.
- Evolution and Co-Evolution: The evolution mechanism Ω drives state transitions by specifying what evolves, when evolution occurs, how variants are generated, where evolution happens, and how quality is evaluated.Co-evolution requires at least two evolving units to jointly adapt and continually reshape each other’s subsequent evolution.
- Three-Stage Taxonomy: The taxonomy has three stages that expand what the system can evolve while progressively removing human-engineered constraints.The stages are Agent–Agent Co-Evolution, Agent–Environment Co-Evolution, and Meta Co-Evolution.
- Stage 1 — Agent–Agent Co-Evolution: Stage 1 allows agents to adapt to dynamic peers, changing one another’s challenges and opportunities while potentially reorganizing the collective structure Π.This replaces learning against a static counterpart with a coupled evolutionary process within the agent collective.
- Stage 2 — Agent–Environment Co-Evolution: Stage 2 extends co-evolution to the environment, allowing agent behavior to reshape interaction conditions that subsequently shape the agents.A fixed environment limits the experiences and evolutionary pressures agents can encounter.
- Stage 3 — Meta Co-Evolution: Stage 3 lets the lower-level co-evolving system revise its evolution mechanism through a self-generated process Γt, providing a pathway toward open-endedness.The proposed characterization includes continuous novelty, Ωt+1 = Ωt, unbounded divergence, and adaptive capability H satisfying limt→∞H(St, Ωt) = ∞.
3 Agent–Agent Co-Evolution
Agent–Agent Co-Evolution describes improvement driven by other evolving agents, organized into adversarial, collaborative, and organizational adaptation. These patterns range from opposing objectives and shared goals to jointly evolving roles and interaction structures.
- 3 Agent–Agent Co-Evolution: The stage centers on agents whose primary source of improvement is the other evolving agents around them, spanning adversarial, collaborative, and organizational patterns.Adversarial and collaborative co-evolution are distinguished by goal relations, while organizational evolution adapts roles and interaction structures.
- Adversarial Co-Evolution: Adversarial co-evolution uses opposing objectives, with each agent’s progress increasing the difficulty faced by its opponent.The literature is organized by pressure scaling from pairwise competition to pressure from multiple adversarial agents.
- Adversarial Co-Evolution: Adversarial systems extend from GAN-style generator–discriminator improvement and disturbance-focused control to LLM attackers and defenders that continually drive safety improvements.Examples include RARL, ACE-Safety’s Monte Carlo tree search, AdvGRPO’s dense multi-channel rewards, and methods that generate new attack types.
- Collaborative Co-Evolution: Collaborative co-evolution uses shared goals, either through parallel cooperation among equivalent agents or through distinct roles with different improvement paths.Role-differentiated systems commonly use produce-and-revise loops, process feedback, verifier bootstrapping, and role-specific credit assignment.
- Evolving Agent Organizations: Organizational co-evolution jointly adapts agents with role assignments and interaction structures, using behavior and execution traces to discover roles, update skills, and restructure teams.R3DM co-adapts roles with agent policies, while SkillMAS jointly updates skills and team structure when traces reveal a mismatch.
4 Agent–Environment Co-Evolution
Agent–Environment Co-Evolution studies systems in which adaptive environments reshape the tasks, feedback, or worlds agents encounter as agents improve. Existing methods therefore evolve task spaces, evaluators, executable environments, and world models to drive further agent evolution.
- Section scope: Adaptive environments reshape agents’ tasks, feedback, or interaction worlds, with environmental change serving as the primary driver of further agent evolution.The stage is organized around how the environment changes what agents solve, how their outputs are evaluated, or where they act.
- Task-space co-evolution: Task-space co-evolution either selects tasks from an existing pool or generates new specifications near the agent’s current competence frontier.Selection methods adapt exposure or difficulty, while generative methods create tasks from goals, failures, interactive behavior, or verification requirements.
- Feedback co-evolution: Feedback co-evolution adapts rewards, evaluators, or diagnostics from preference comparisons, task outcomes, critic agreement, and consistency constraints.Methods reuse or relabel trajectories, seek uncertain comparisons, revise reward functions from failures and returns, and constrain evaluator updates using agreement or outcome consistency.
- World-space co-evolution: World-space co-evolution constructs increasingly challenging executable environments by targeting agent weaknesses, adapting rules or layouts, and evolving levels, controllers, or population incentives.Examples span gridworlds, continuous control, robotics, embodied simulation, games, open-ended worlds, tool-use environments, economic incentives, and social norms.
- World-model co-evolution: World-model methods replace real environments with co-evolving virtual settings that synthesize rollouts, predict future feedback, or generate transitions during agent training.WebEvolver, COMAP, and DreamGym respectively use web simulation, textual feedback prediction, and experience-model transition synthesis.
5 Meta Co-Evolution
Meta co-evolution makes the evolution mechanism itself adaptive, potentially overcoming performance plateaus and enabling open-ended systems with continuous novelty and unbounded divergence. The section frames this progression through adaptive decisions and contrasts single-entity precursors with systems that coevolve agents and evaluators.
- Meta Co-Evolution: Meta co-evolution may overcome the plateau reached by Stage 1 and Stage 2 by allowing the evolution mechanism itself to change.Stage 1 and Stage 2 improve performance across most settings, but their gains diminish as evolution approaches a plateau.
- Meta Co-Evolution: Open-ended meta co-evolution requires continuous novelty and unbounded divergence in the system’s adaptive capacity.The passage formalizes these requirements as Ωt+1̸ = Ωt and limt→∞H(St, Ωt) = ∞.
- Evolution Mechanism: The evolution mechanism can be decomposed into adaptive decisions about what, when, and how to evolve.These decisions select adaptation targets, trigger updates after failures or plateaus, and produce variants through training, revision, or structural generation and editing.
- Related Work: Most related methods are single-entity precursors that evolve prompts, self-modification machinery, memory architectures, or update choices without a lower-level co-evolving system.PromptBreeder, Gödel Agent, HyperAgents, MemEvolve, and SIA make the evolution mechanism evolvable but remain limited in this respect.
- Related Work: RQGM extends beyond these precursors by coevolving task agents and evaluators while using a meta-agent and joint feedback to guide later evolution.This combines coevolution of task agents and evaluators with feedback-driven guidance of subsequent evolution.
6 Challenges and Future Directions
The section identifies dynamic evaluation, scaling co-evolution beyond local loops, and safety and governance as key challenges for increasingly autonomous agentic systems. Future work must evaluate adaptation over time, coordinate changes across agents and environments, and preserve human control.
- Dynamic evaluation: Dynamic evaluation must move beyond benchmarks that mainly measure an agent’s final capabilities across domains.Existing benchmarks cover tool use, web browsing, software engineering, computer use, and multi-agent interaction.
- Scaling co-evolution: Scaling co-evolution requires extending local attacker–defender, policy–reward, and agent–task loops to systems where agents, harnesses, and environments change together.The central challenge is deciding which components should change and how their updates influence one another.
- Safety and governance: As co-evolution becomes more autonomous, evolving agents may develop behaviors that exceed human understanding and are difficult to monitor, weakening human control.These behaviors can include attack strategies, tool-use patterns, communication protocols, and organizational behaviors.
- Safety and governance: Meta co-evolution sharpens safety risks because systems may alter which behaviors are rewarded and preserved.Future governance must address this possibility while keeping increasingly autonomous systems controllable.
7 Conclusion
The survey frames co-evolution as a progressive expansion from agent adaptation to environmental and evolutionary-mechanism adaptation, reducing human intervention over time.
- 7 Conclusion: Co-evolution occurs when multiple components impose evolutionary pressure on one another.
- 7 Conclusion: The taxonomy expands its adaptive boundary from agents alone to agents and environments, then to the mechanism governing their joint evolution.
- 7 Conclusion: This progression reflects a gradual reduction of human intervention in agentic-system evolution.
Limitations
The survey finds that meta co-evolution remains underdeveloped, while safety and governance for increasingly autonomous co-evolution are identified but not operationalized through concrete safeguards or protocols.
- Meta Co-Evolution: Meta co-evolution remains at an early stage, with limited work meeting the Stage 3 definition.The discussion therefore relies substantially on single-entity meta-evolution as a precursor.
- Meta Co-Evolution: Existing single-entity meta-evolution shows that evolution mechanisms can themselves evolve, but not alongside a lower-level co-evolving system.
- Safety and Governance: Safety, monitoring, and human oversight are identified as first-order concerns, but remain unoperationalized.The survey does not develop concrete safeguards or protocols for increasingly autonomous co-evolution.
- Safety and Governance: Co-evolving systems present specific failure modes, including evaluator exploitation, partner overfitting, and diversity collapse.These risks are treated as desiderata rather than addressed through concrete safeguards or protocols.
Ethics Statement
The paper surveys how agents and environments adaptively reshape one another toward self-directed evolution. It reports that reviewed works are cited, materials are believed publicly accessible or research-permitted, and no additional dataset curation or human annotation was conducted, so the authors identify no ethical concerns.
- The survey examines co-evolution in agentic systems as agents and environments adaptively reshape one another toward self-directed evolution.
- All reviewed research works are properly cited, and the referenced materials are believed publicly accessible or licensed for research review.
- The authors conducted no additional dataset curation or human annotation and therefore believe the paper raises no ethical concerns.
Appendices · A Distinguishing Co-Evolution from Adjacent Concepts · B Comparison with Related Surveys
The appendices distinguish co-evolution from related concepts by requiring multiple persistent adaptive units that reshape one another, and compare this criterion with adjacent survey taxonomies. They also clarify that co-evolution differs from task-execution design, candidate selection, single-learner improvement, and open-endedness.
- A Distinguishing Co-Evolution from Adjacent Concepts: Co-evolution requires multiple adaptive units whose persistent changes continue shaping one another, unlike multi-agent interaction without agent updates.Interaction becomes co-evolution only when changes persist across agents and alter their subsequent adaptation.
- A Distinguishing Co-Evolution from Adjacent Concepts: Agent loops and harness engineering concern task execution or runtime design, becoming evolution only when the model, harness, or both are updated and retained across runs.Loop engineering specifies goals, feedback, verification, and stopping rules, while harness engineering specifies prompts, memory, tools, skills, and workflows.
- A Distinguishing Co-Evolution from Adjacent Concepts: Evolutionary optimization selects evaluated candidate prompts, programs, or designs, whereas co-evolution requires persistent adaptive units responding to one another.Continual learning instead accumulates changes within one continuing learner, and its learning procedure may remain fixed.
- A Distinguishing Co-Evolution from Adjacent Concepts: Self-evolution, continual learning, self-play, and harness evolution can involve persistent change while still remaining within one evolving agent or learner.Co-evolution additionally requires a distinct adaptive unit that reshapes the agent’s later evolution and is itself reshaped.
- A Distinguishing Co-Evolution from Adjacent Concepts: Meta Co-Evolution requires evolution-mechanism changes to be driven by and then reshape a lower-level co-evolving system.Meta-evolution can modify what, when, how, and where to evolve and how changes are evaluated, even if its controller remains fixed.
- A Distinguishing Co-Evolution from Adjacent Concepts: Iterative and recursive self-improvement may operate within one system, while open-endedness describes a possible outcome rather than a specific evolutionary form.Co-evolution therefore does not necessarily follow from recursive improvement and does not automatically guarantee open-endedness.
- B Comparison with Related Surveys: Existing surveys organize literature around multiagent systems, harnesses, self-evolution, recursive improvement, or environment scaling, leaving co-evolution scattered or secondary.This survey instead uses co-evolution as its central inclusion criterion and locates it across agents, environments, and evolution mechanisms.
C Construction of the Cross-Paper Evidence Figure
Figure 4 constructs cross-paper evidence through within-paper matched comparisons and normalized evolution trajectories, rather than pooling raw scores across heterogeneous studies. Its panels summarize static-versus-evolving differences, paired improvements, and paper-level trajectory means while preserving cross-paper heterogeneity.
- Matched comparisons in Panels A and B: Included papers must report static or frozen and evolving counterparts under the same backbone and evaluation setting, using a controlled evolution-only variant when alternatives exist.Settings lacking results for both conditions are excluded.
- Matched comparisons in Panels A and B: Panel A averages direction-aligned scores separately for static and evolving conditions, but its endpoints support only within-paper comparisons and are not cross-paper absolute scores.Higher values consistently indicate better performance after direction alignment.
- Matched comparisons in Panels A and B: Panel B plots the number of matched settings against mean paired improvement, with bubble area showing the share of positive improvements and color denoting taxonomy stage.Unmatched or aggregate-only results are not counted as individual evaluation settings.
- Normalized trajectories in Panel C: Panel C normalizes reported performance trajectories over evolution progress, direction-aligns lower-is-better metrics, excludes trajectories without positive improvement, and interpolates them onto a common grid.Values use explicit labels when available and approximate digitization otherwise.
- Normalized trajectories in Panel C: Multiple trajectories within a paper are averaged into one equally weighted paper-level curve; the solid line is their mean and shading is one sample standard deviation reflecting cross-paper heterogeneity.The variability band represents heterogeneity rather than experimental uncertainty or a confidence interval.