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

Abstract 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 humandesigned paths.

Introduction. Recent advances in AI have shifted attention from isolated models to agentic systems, which can autonomously interact with other agents and external environments (Guo et al., 2024; Plaat et al., 2025; Luo et al., 2025). They may be augmented with a harness (Ning et al., 2026; Yao et al., 2026b), including tools (Xu et al., 2025b), memory (Du, 2026), and skills (Zhou et al., 2026). As agentic systems become increasingly complex, a central question is how they can continue to improve after deployment (Gao et al., 2026). Self-evolution is a major paradigm for continual improvement in agentic systems, where an agent persistently updates itself from experience, feedback, and failures without human intervention. One important form is single-entity selfevolution, where updates occur within one agent, such as refining its model backbone (Zhai et al., 2025a; Wu et al., 2025), updating memory (Suzgun et al., 2026), or improving skills (Xiao et al., 2026). However, it can remain bounded by fixed external conditions, as illustrated in Figure 1.

Discussion / Conclusion. Dynamic evaluation. Existing benchmarks mainly measure an agent’s final capabilities across domains such as tool use (Yao et al., 2024; Lu et al., 2025b; Liu et al., 2026b,c,d), web browsing (Zhou et al., 2024a; Wei et al., 2025), software engineering (Jimenez et al., 2024; Miserendino et al., 2025), computer use (Trivedi et al., 2024; Xie et al., 2024; Merrill et al., 2026), and multi-agent interaction (Zhou et al., 2024b; Mou et al., 2025; Zhu et al., 2025a). PostTrain- Bench (Rank et al., 2026) takes a step toward evolution-aware evaluation by testing whether an agent can autonomously post-train a base LLM. However, evaluating co-evolution requires going further: it must determine whether all evolving components improve, whether their gains transfer to unseen partners and environments, and how each component contributes to the joint progress.