When an AI tunes its own setup, the feedback signal that feels most reliable — task success — turns out to be the most misleading.

This explores what actually moves the needle when an agent evolves its own harness — the prompts, tools, and objectives that scaffold its behavior — rather than assuming any feedback loop that produces changes is producing improvement. The corpus's sharpest warning is that the most seductive signal, raw task success, is also the most misleading: harness evolution is itself a search loop, so gains reported against it are confounded with sheer search effort Are harness evolution gains really from better design?. The feedback that matters is the delta *beyond* a matched search budget, measured on held-out tasks — otherwise you're rewarding memorization, not design. This is borne out by inspection: most evolved edits just cache task-specific fixes the agent could have rediscovered in a single rollout, so the signal to distrust is one that improves already-solvable tasks rather than converting failures into successes Do harness edits learn strategies or just memorize fixes?.

If outcome scores are noisy, the corpus points to a more reliable signal hiding underneath them: persistence through the loop. Across 17 frontier models on long-horizon optimization tasks, the dominant predictor of success wasn't the quality of any single edit but whether the agent kept running the benchmark-edit-incorporate cycle instead of terminating early or burning budget unproductively What predicts success in ultra-long-horizon agent tasks?. So during search, a signal worth tracking is whether feedback is being *incorporated* iteration over iteration — not just whether any one iteration scored well. Feeding that loop requires dense, per-step signals rather than a lone scalar, which is why decoupling evaluation into separate benchmark, harness, and environment components matters: it surfaces reward-hacking and failure modes in the trajectory that a single success number hides entirely How can separating evaluation components make reward-hacking visible?.

The deepest constraint is that self-generated feedback alone runs out. Pure self-improvement stalls on the generation-verification gap and reward hacking, and the methods that actually work smuggle in external anchors — past model versions, third-party judges, user corrections, tool feedback Can models reliably improve themselves without external feedback?. Harness evolution is a self-improvement loop, so the same rule applies: the feedback that keeps it honest comes from outside the agent's own judgment. Tree search offers one way to manufacture such a signal without human labels — MCTS outcomes plus critic models yield dense process-level rewards that rank solution paths by whether they actually succeed Can tree search replace human feedback in LLM training?.

There's also a wrinkle in *who* the feedback is for. The capacity to produce useful harness edits is flat across model tiers, but the capacity to benefit from them follows an inverted U — weak models never invoke the harness, strong ones struggle to follow their own instructions faithfully Do stronger models always evolve harnesses better?. So a signal that works for a mid-tier planner may be inert for others, and how the harness is *represented* changes what feedback is even actionable: reorganizing it around runtime behavior lets a weaker planner localize as well as a stronger model Can explicit behavior maps help weaker planners compete with stronger models?.

Finally, the frontier framing turns feedback from an input into something the search itself produces. Evolutionary inference-time search beats Best-of-N and sequential revision partly because an island model sustains population diversity — the signal it needs is not just 'is this good' but 'is this *different*,' since entropy-collapsed candidates can't be recombined into anything new Can evolutionary search beat sampling and revision at inference time?. That's why training upstream for diversity beats scalar optimization when a model feeds into search Should training maximize diversity when models feed into search?. And a bi-level agent can close the loop entirely by evolving its own objective functions — compiling natural-language goals into executable scoring code — so the feedback signal becomes a first-class thing the search designs rather than inherits Can agents evolve their own objectives during search?. The through-line: the feedback that matters most is external, budget-matched, diversity-aware, and measured by conversion of failures — not the raw task score the loop is happiest to optimize.