Can agents evolve beyond the constraints humans engineer?

A survey of co-evolution in agentic systems starts from a limit on the single-entity picture of self-improvement. The introduction defines self-evolution as an agent persistently updating itself "from experience, feedback, and failures without human intervention," and names the single-entity form: refining a model backbone, updating memory, improving skills. It then says this form "can remain bounded by fixed external conditions." The abstract makes the limit specific, calling single-entity self-evolution "often bounded by a static learning context, such as fixed tasks and feedback." Co-evolution is the proposed way out: a multi-component form of self-evolution in which "multiple agents and their environment impose adaptive pressure on one another."

The survey's organizing device is a "progressive three-stage taxonomy that traces how the system gradually sheds human-engineered constraints." Agent–Agent Co-Evolution covers adaptation through dynamic peers, whether adversarial, collaborative, or organizational. Agent–Environment Co-Evolution extends the loop to "adaptive tasks, feedback, and interaction spaces that change with the agents." Meta Co-Evolution asks whether "the evolution mechanism itself" can be made evolvable. Read in order, each stage removes one more thing a human would otherwise fix: first the partners, then the tasks and feedback, finally the update rule. The discussion adds an evaluation requirement. Existing benchmarks "mainly measure an agent's final capabilities," and PostTrainBench is described as only a step toward evolution-aware evaluation. Evaluating co-evolution must establish whether all evolving components improve, whether gains transfer to unseen partners and environments, and how each component contributes to joint progress.

This sits on a different axis from Do self-improving agents really split into two distinct loops?, which sorts self-improvement by what gets updated, weights or scaffold. The co-evolution taxonomy sorts it by what supplies the pressure and how much of the loop stays human-designed, so a system could be placed on both axes. Can evaluators improve alongside the agents they score? reads as a case of the second stage, since a learned evaluator that improves alongside its agents is adaptive feedback in the survey's sense, though the excerpt does not say which papers it places where. The survey's dynamic-evaluation requirement also echoes Does a single benchmark score actually predict agent readiness?, with per-component contribution and transfer to new partners as further axes. My inference is that the third stage sharpens the concern in How can agent self-evolution be made safe and auditable?: if the evolution mechanism itself changes, that mechanism needs lineage and rollback too.

The excerpt is a survey abstract plus one introduction passage and one discussion passage, so much is not established. It gives no count of surveyed papers, no comparison showing co-evolution outperforming single-entity self-evolution, and no evidence that the three stages form a developmental sequence rather than an organizing convenience. The open challenges are only named: evaluation, scaling "across multiple components," and keeping autonomous evolutionary processes "safe and controllable." The defensible takeaway is limited to a map of the design space and a stated gap in evaluation. The excerpt does not show that removing human-engineered constraints improves outcomes.