When an AI rewrites its own toolkit based only on pass/fail scores, does it get smarter or just memorize the test?
Do task-level outcomes provide sufficient supervision for harness evolution?
This explores whether pass/fail results on tasks are enough of a signal to steer an agent as it rewrites its own scaffolding (prompts, tools, memory rules), or whether that signal mostly teaches it to game the tasks it has already seen.
This explores whether pass/fail results on tasks are enough of a signal to steer an agent as it rewrites its own scaffolding (prompts, tools, memory rules), or whether that signal mostly teaches it to game the tasks it has already seen. The corpus leans toward: enough to pick a winner, not enough to teach what's worth keeping. An outcome score says an edit helped on these tasks. It doesn't say why, and that gap is where memorization creeps in. One analysis of evolved harnesses found the edits were sensible and well-motivated, yet most just persisted fixes an agent could have rediscovered in a single rollout. They cached what was already within reach instead of turning hard failures into successes (Do harness edits learn reusable strategies or memorize task fixes?). Related work finds that recursive scaffold edits can memorize the tasks they were evolved on, so in-distribution gains shrink out of distribution. The proposed fix is to constrain both how edits are proposed and how they are selected, favoring reusable mechanisms over benchmark-specific ones (Does harness self-improvement memorize tasks instead of learning broadly?).
The outcome is also a blurrier signal than it looks. A model's ability to write useful harness edits is roughly flat across capability tiers, but its ability to benefit from them follows an inverted U. Weak models fail to invoke the harness, and very strong ones struggle to follow its instructions faithfully (Do stronger models always evolve harnesses better?). So a failed run could mean a bad edit or a model that can't use a good one, and the score can't tell them apart. There is a budget confound too. Task outcomes are the same feedback that plain test-time search uses, so a harness only earns credit for gains beyond what matched-budget search would have found (How should we measure gains from automatic harness evolution?).
What seems to help is changing the evidence, not just the score. One approach evolves harness modules separately, using contrastive trajectories (comparing what differed between runs) on benchmark-disjoint data. It aggregates evidence across tasks before touching any component, and the gains hold on unseen tasks and domains (Can harness modules improve separately from benchmark data?). Scale across environments helps in a similar way. An auto-research loop run over many environments surfaced four mechanisms (action execution, context compaction, observation handling, delegated reading) that cut token traffic by 44.7–49.0% at comparable performance. Those are efficiency gains, not accuracy gains (Can agent harnesses be automatically optimized across many environments?). The edits that transfer also tend to be structural. A stronger model nearly doubled a weaker one's Theory-of-Mind performance mainly by moving unstable reasoning into deterministic code and routing (Can a stronger model lift a weaker one at test time without retraining?). Reorganizing a repository around a behavior-to-code map let weaker planners match stronger models at localizing code (Can explicit behavior maps help weaker planners compete with stronger models?).
Outcome-driven evolution can still generalize when candidates are checked on tasks they weren't selected on. AIDE85, evolved through seven accepted rewrites over eight days, matched or beat its human-built counterpart on four held-out benchmarks (Does automated evolution match human-built agent performance?). One of those was physics-based weather forecasting, outside the selection distribution (Do AIDE2's improvements transfer to unseen tasks?). So outcomes work as the judge when the gate is held-out, but they are weak as the teacher. The corpus has no head-to-head between outcome-only and richer supervision, so that split is an inference from these pieces. One survey adds that self-improvement in a static setting stalls, and the feedback signal itself is a component worth evolving (Can agents evolve beyond the constraints humans engineer?).
Sources 11 notes
Analysis of evolved harness trajectories shows rational, well-motivated edits across prompt and tool layers, but most persist fixes an agent could rediscover in a single rollout. Gains remain limited because memorized shortcuts cache what's already within reach rather than converting hard failures into successes.
Recursive edits to agent scaffolding can memorize evolve-set tasks, causing in-distribution gains to shrink out-of-distribution. RRSI addresses this by constraining the proposal and selection stages to favor reusable mechanisms over benchmark-specific ones.
Model capability to produce useful harness edits stays constant across tiers, but capacity to actually benefit from those edits follows an inverted U-shape, peaking in mid-tier models. Weak models fail to invoke harnesses; strong models struggle with faithful instruction-following.
Automatic harness evolution must be compared against task-level test-time search under equal feedback and inference budgets. Only gains beyond what matched search achieves are attributable to the harness design itself, not just more computation.
ModularRSI evolves harness modules independently using contrastive trajectories on benchmark-disjoint data, showing consistent gains across unseen tasks and domains. The approach isolates mechanism-level improvements from task-specific adaptation by aggregating evidence across tasks before updating components.
Show all 11 sources
Cross-environment harness optimization yielded four mechanisms (action execution, context compaction, observation handling, delegated reading) that reduced token traffic by 44.7–49.0% while maintaining comparable performance on a 51-task benchmark, suggesting harness-level gains are orthogonal to model improvements.
A stronger model built inference-time harnesses that nearly doubled weaker model performance on Theory-of-Mind benchmarks without retraining, primarily by moving unstable reasoning into deterministic code and task-specific routing rather than encouraging extended reasoning.
A behavior-to-code mapping representation improved win rates by 10–19 points while reducing planner tokens by 8–13%. Weaker planners using this mapping matched stronger models' code localization across all precision and recall metrics.
AIDE85, evolved through seven accepted rewrites in 8 days, equals or surpasses AIDEhuman on four held-out benchmarks spanning in- and out-of-distribution tasks including weather forecasting. The result shows automated design iteration can match human-driven R&D on generalization.
The paper reports that AIDE2's improvements transfer to four held-out benchmarks spanning machine learning, algorithm engineering, and physics-based weather forecasting—the last being outside the selection task distribution. This demonstrates transferable gains beyond overfitting to the selection set.
A survey framework organizes co-evolving systems into three stages that progressively remove human engineering: dynamic peers first, then adaptive environments and feedback, finally the evolution mechanism itself. Single-entity self-improvement stalls in static contexts; co-evolution supplies adaptive pressure across multiple components.
Papers this line draws on 8
The research behind the notes this line reads — ranked by how closely each paper relates.
- Rethinking the Evaluation of Harness Evolution for Agents
- Harness Updating Is Not Harness Benefit: Disentangling Evolution Capabilities in Self-Evolving LLM Agents
- DarwinX: Evolving Agent Harnesses Through Natural Selection
- ModularRSI: Modular and Generalizable Recursive Harness Self-Improvement
- HarnessDev: Can LLMs Create and Evolve Their Own Agent Harness?
- Ouroboros: A Self-Developing Frontier Coding Agent with Reviewed Core Evolution
- Harness Handbook: Making Evolving Agent Harnesses Readable,Navigable, and Editable
- Co-Evolution in Agentic Systems: Toward Self-Directed Evolution Beyond Human Design