Co-Evolution in Agentic Systems: Toward Self-Directed Evolution Beyond Human Design

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.

The paper defines an agentic system S = (A, E) consisting of an agent collective A and an environment E. An agent is a system unit that can autonomously interact with other agents or the environment, denoted by ai = (mi, hi): a model backbone mi and a harness hi, such as memory, tools, skills, prompts, or workflows. Agents are structured by Π, which encodes their roles, communication topology, and division of labor. The environment E is everything external to A that agents act upon and receive feedback from.

Co-evolution is defined as a coupled form of self-evolution in which at least two evolving units jointly adapt and continually reshape each other’s further evolution, rather than merely exchanging information or interacting. It requires that both units change and that they exert evolutionary pressure on each other.

The three-stage taxonomy is structured by the expanding scope of what a system is allowed to evolve. Stage 1 — Agent–Agent Co-Evolution captures mutual adaptation among evolving peers within a fixed environment, including changes to their model backbones, harness components, and organizational structures. This stage is organized into three patterns: adversarial and collaborative co-evolution distinguished by agents’ goal relations, and evolving agent organizations that further adapt agents’ roles and interaction structures. Adversarial agents co-evolve through opposing objectives, where each side is rewarded for defeating the other, and can involve pairwise pressure between two evolving sides or multi-source pressure across many adversarial agents. Collaborative agents co-evolve through shared goals, distinguished into parallel collaboration, where equivalent agents either have no roles or can take any role, and role-differentiated collaboration, where fixed and distinct roles give agents different paths of improvement. Evolving agent organizations go beyond fixed roles, allowing agents to co-evolve with the organization, including role assignments and interaction structure.

Stage 2 — Agent–Environment Co-Evolution extends adaptation to environmental components, including tasks, feedback, and interaction spaces. In task-space co-evolution, the environment adapts what problem the agent solves next, taking two forms: selecting tasks from an existing pool and generating new task specifications. In feedback-space co-evolution, the environment co-evolves in how it evaluates, rewards, or diagnoses agent behavior, driven by preference comparisons, task outcomes, or consistency constraints. In interaction-space co-evolution, the environment adapts where the agent acts in, either constructing an increasingly challenging executable world or learning a world model to replace the real one.

Stage 3 — Meta Co-Evolution captures an emerging stage in which the evolution mechanism itself becomes adaptive, allowing the system to revise what, when, how, and where to evolve, as well as how to evaluate. The evolution mechanism omega specifies what can evolve, when evolution is triggered, how variants are generated or updated, where evolution takes place, and how evolution quality is evaluated. Meta co-evolution is defined as a stage where the lower-level co-evolving system further revises its evolution mechanism through a self-generated revision process Γt: omegat+1 = Γt(S t, omegat, τ t), S t+1 = omegat+1(S t, τ t). This recursion provides a pathway toward open-endedness, where the system continues to generate meaningful novelty rather than converging to a fixed endpoint, characterized by continuous novelty, omegat+1 ≠ omegat, and unbounded divergence, with adaptive capability H satisfying limt→∞ H(S t, omegat) = ∞.

The paper decomposes the evolution mechanism into five adaptive decisions: (1) what to evolve selects the adaptation target, such as an agent backbone, harness, or organization, or environment artifacts like tasks, rewards, and worlds; (2) when to evolve triggers updates after failures, plateaus, or distribution shifts; (3) how to evolve produces variants through backbone training, harness revision, or structural generation and editing; (4) where to evolve covers both the domain of evolution, such as tool use, web, or robotics, and the setting it runs in, such as sandboxes, simulated or real environments; and (5) how to evaluate judges evolution quality by performance, novelty, safety, or robustness, shaping the direction of later evolution.

The paper discusses open challenges in dynamic evaluation, scaling co-evolution, and safety and governance. For dynamic evaluation, it notes that evaluating co-evolution requires determining whether all evolving components improve, whether their gains transfer to unseen partners and environments, and how each component contributes to the joint progress. Since higher task success can hide exploitative behavior, such systems may fail through evaluator exploitation, partner overfitting, or diversity collapse. For scaling co-evolution, the challenge is not to make more components adaptive, but to decide which should change, how their updates influence one another, and how to keep the co-evolution pressure productive rather than unstable or dominated by one component. For safety and governance, as co-evolution becomes more autonomous, humans may lose control over the system’s behavior, and future work needs governance that keeps open-ended evolution auditable and interruptible, including sandboxed deployment, continuous monitoring, rollback to verified states, and human intervention points.

The paper concludes that future progress lies not in stronger static agents but in agents that continually improve through co-evolution, supported by evaluation and governance that keep the process reliable and controllable. The limitations noted are that meta co-evolution is still at an early stage with only limited work meeting the definition of Stage 3, and that safety and governance are not operationalized but only treated at the level of desiderata.

Improvements to AI systems:

1. Implement adaptive multi-agent training loops where agents are explicitly coupled so that changes in one agent’s policy, memory, or tools directly impose selection pressure on peers (adversarial or collaborative). The improved system can self-organize into specialized roles (e.g., generator vs. critic, explorer vs. exploiter) and continuously rebalance its division of labor without human retuning.

2. Add co-evolving task generators that propose new problems based on the current capability frontier of the agent collective. The improved system can automatically generate curricula of tasks that are neither too easy nor too hard, preventing saturation and promoting open-ended skill acquisition.

3. Enable feedback-space co-evolution where the reward function or evaluation criteria themselves adapt based on task outcomes and consistency constraints. The improved system can detect when its current reward is being gamed, revise the reward to penalize exploitation, and maintain alignment with underlying goals even as behavior becomes more complex.

4. Build interaction-space co-evolution that constructs increasingly challenging simulated worlds or learns a world model to replace the real environment. The improved system can train in sandboxed, procedurally generated environments that grow in complexity, then transfer robustly to real-world settings with less human supervision.

5. Introduce meta-evolution mechanisms that allow the system to revise its own evolution rules (what to evolve, when, how, where, and how to evaluate). The improved system can detect plateaus in performance, switch its adaptation target (e.g., from agent weights to tool design), change its evaluation metric to prioritize novelty or safety, and thereby avoid converging to local optima.

6. Add dynamic evaluation and safety monitors that track per-component improvement, transfer to unseen partners/environments, and flag exploitative behavior (e.g., reward hacking, partner overfitting, diversity collapse). The improved system can roll back to verified states, pause evolution when instability is detected, and request human intervention at predefined checkpoints, ensuring controllability during autonomous co-evolution.

What the improved AI system can do:

• Self-improve indefinitely across multiple interacting agents and environmental components, without human-designed fixed curricula or static reward functions.

• Adapt its own evolution strategy—choosing which part of the system to change, when to trigger an update, and how to evaluate progress—leading to open-ended novelty rather than diminishing returns.

• Maintain safety and reliability during autonomous evolution by continuously monitoring for exploitative or unstable behaviors, and by keeping human-interruptible checkpoints and rollback capabilities.

• Transfer learned capabilities to new tasks, partners, and environments because co-evolution forces robustness against varied pressures, not just overfitting to a single static benchmark.