If AI agents evolve side by side, pushing and testing each other, do they eventually design themselves — no human engineer needed?

This explores whether letting AI agents evolve alongside each other, as well as alongside their environments and goals, can take over design work that human engineers would otherwise do by hand. The corpus mostly says yes, but in stages. One survey framework lays out three steps. First, agents get dynamic peers. Next, the environment and the feedback adapt too. Last, the evolution mechanism itself starts to evolve Can agents evolve beyond the constraints humans engineer?. The core argument is that a single agent improving itself in a fixed setting eventually stalls, because nothing new pushes back on it. Peers supply that push. Peer co-evolution is the first rung of that ladder, not the whole climb.

The most concrete evidence that peers matter comes from Group-Evolving Agents. Agents that pooled code patches and execution traces within each generation beat isolated self-evolving lineages by 14–20 points. Five of the eight key tool improvements came from a different parent than the one that used them, so the gain came from the sharing itself, not just from running more search Does sharing experience across agents beat isolated evolution?. SkillClaw does something similar at the level of users: it gathers interaction histories across many people and lets an automated evolver refine shared skills that a human would otherwise have to curate How can agent systems share learned skills across users?. In game settings, agents trained against a varied mix of co-players learn to cooperate without anyone hardcoding cooperative rules. Being exposed to exploitation by each other pushes them toward cooperation Can agents learn cooperation by adapting to diverse partners?.

A quieter lesson is that "peers" don't have to be other agents. POET pairs agents with environments that evolve alongside them and lets solutions move between problems. It solved obstacle courses that neither direct optimization nor hand-designed curricula could, and that movement between problems turned out to be essential Can paired environment and agent optimization unlock unsolvable challenges?. SAGA goes a step further and evolves the objective itself. An outer loop proposes new goals and compiles them into scoring code, so the goal stops being a fixed human input Can agents evolve their own objectives during search?. This is where human design really drops out: not when agents talk to each other, but when the environment and the goal stop being fixed by people.

Does the result actually match human work? In at least one case, yes. An agent rewritten through seven automated revisions over eight days equaled or beat its human-engineered counterpart on four held-out benchmarks, including weather forecasting Does automated evolution match human-built agent performance?. Most of this progress happens in the cheap "fast loop" of prompts, memory, and tools, not in model weights, because scaffold changes are cheap and easy to undo Do self-improving agents really split into two distinct loops?. There's a catch, though. Models of every size are about equally good at proposing improvements to their harness, but mid-tier models benefit most from them. Weak models don't use the changes, and strong ones don't follow them faithfully Do stronger models always evolve harnesses better?. So co-evolution doesn't automatically pay off more as models get more capable.

The corpus doesn't directly measure how much human design effort any of these systems saves, so "reduces reliance" holds up in performance terms, not in a counted tally of human work. Humans still choose the starting setup, what gets shared, and the outer loop, at least until the third stage of the survey's ladder.