When an AI rewrites its own instructions after failing, does it learn real lessons or just memorize answers to old problems?
Can harness edits distill reusable strategies or mostly memorize task-specific fixes?
This explores whether automated edits to an agent's scaffolding (prompts, tools, memory, skill files) capture general strategies that transfer to new tasks, or mostly cache fixes for the specific problems the agent already saw.
This explores whether automated edits to an agent's scaffolding (its prompts, tools, and instructions) capture general strategies that carry over to new tasks, or mostly cache fixes for the specific problems the agent already saw. The corpus leans toward memorization by default, but it also shows that the outcome can be steered.
The bleakest evidence comes from a close reading of evolved harness trajectories. The edits look sensible one at a time, across both prompt and tool layers. But most of them persist fixes an agent could have found again in a single rollout (Do harness edits learn reusable strategies or memorize task fixes?). Gains stay small because the edits cache what was already within reach and don't turn hard failures into successes. The measured overfitting matches this: gains on the tasks used to evolve the harness shrink on tasks outside them (Does harness self-improvement memorize tasks instead of learning broadly?).
The fix looks like constraining the process, not hoping for better edits. One approach evolves harness modules separately, using contrastive trajectories on data that is disjoint from the benchmark. It gathers evidence across many tasks before touching a component, so a mechanism has to pay off broadly to survive (Can harness modules improve separately from benchmark data?). On the skill-document side, capping how much text can change per step, gating edits on held-out validation, and keeping a buffer of rejected edits as negative feedback all beat letting an agent freely rewrite its own instructions (Does constraining edits make skill learning more stable?, Can skill documents be optimized like neural network weights?). Held-out gating makes an edit prove itself on data it wasn't written for, and that same check is the one memorized fixes fail.
Some work shows real distillation happening. Optimizing across many environments produced four mechanisms (action execution, context compaction, observation handling, delegated reading) that cut token traffic by 44.7–49.0% at comparable performance (Can agent harnesses be automatically optimized across many environments?). Agent Workflow Memory does something similar at the strategy level. It extracts sub-task routines that are finer than whole tasks and strips out example-specific values. Its gains are largest, 24.6% on Mind2Web and 51.1% on WebArena, exactly when the train-test gap widens (Can agents learn reusable sub-task routines from past experience?). AIDE2's gains also held on four unseen benchmarks, including physics-based weather forecasting outside the selection distribution (Do AIDE2's improvements transfer to unseen tasks?). The common thread is abstraction: reusable edits describe a behavior, not an instance. That may explain why organizing a harness around runtime behavior helped weaker planners match stronger models at code localization (Can explicit behavior maps help weaker planners compete with stronger models?).
The model doing the editing matters too, in a way that's easy to miss. The ability to produce useful harness updates is flat across model tiers, but the ability to benefit from them peaks in mid-tier models. Weak models don't invoke the harness, and strong models follow its instructions less faithfully (Do stronger models always evolve harnesses better?). Building and evolving a harness is also a separate skill from doing the task, and models are poor at keeping useful intermediate updates (Can language models build and maintain their own agent harnesses?). A strong builder can still lift a weaker model by moving unstable reasoning into deterministic code and task-specific routing (Can a stronger model lift a weaker one at test time without retraining?). That last case shows a fix can be task-specific and still worth having.
So the honest answer is that unconstrained evolution mostly memorizes, and the corpus doesn't show that changing. Distillation shows up when the process is built for it: held-out gates, evidence pooled across tasks, and edits pitched at the level of behaviors and routines. The corpus doesn't settle how much of the generalization in the positive results comes from the constraints and how much from the tasks being easier to generalize across.
Sources 12 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.
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.
SkillOpt's ablations show that adding a textual learning-rate budget, held-out validation gate, and rejected-edit buffer (retaining failed edits as negative feedback) produces more stable and generalizable skill improvement than allowing agents to freely rewrite their own instructions.
SkillOpt treats skill documents as trainable external state of frozen agents, using a text-space optimizer with held-out validation gating to accept only edits that improve performance. Across 52 benchmark cells and seven models, the approach matches or exceeds baselines while adding zero inference cost and enabling transfer across models.
Show all 12 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.
Agent Workflow Memory induces sub-task routines at finer granularity than full tasks, abstracts example-specific values, and compounds them hierarchically. This produces 24.6% relative gain on Mind2Web and 51.1% on WebArena, with larger gains as train-test gaps widen.
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 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.
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.
Research shows that LLMs vary sharply in building harnesses across domains, struggle to retain useful intermediate updates during evolution, and produce harnesses whose performance shifts dramatically with different executors—demonstrating that harness quality cannot be inferred from downstream task scores alone.
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.
Papers this line draws on 8
The research behind the notes this line reads — ranked by how closely each paper relates.
- Harness Updating Is Not Harness Benefit: Disentangling Evolution Capabilities in Self-Evolving LLM Agents
- Rethinking the Evaluation of Harness Evolution for Agents
- DarwinX: Evolving Agent Harnesses Through Natural Selection
- HarnessDev: Can LLMs Create and Evolve Their Own Agent Harness?
- Harness Handbook: Making Evolving Agent Harnesses Readable,Navigable, and Editable
- ModularRSI: Modular and Generalizable Recursive Harness Self-Improvement
- RRSI: Regularized Recursive Self-Improvement of Agent Harnesses
- SkillOpt: Executive Strategy for Self-Evolving Agent Skills