SkillClaw: evolve skills from agent feedback

SkillClaw (arXiv:2604.08377) grows an agent's skill pool from the agent's own traffic. By day the agent drains a frozen task list against the current pool. By night one decision per observation changes the pool, and the next day measures what changed.

Evolves	the harness tree: the agent's skill pool
Signal	one report per task, with the day's last report closing the batch
Loss family	none (no weight training)
Package	recipes/skillclaw/
Pipeline	reported feedback, producing trace items
Needs	a Reef process, the pi binary, and Docker for the Harbor tasks. Reef itself needs no GPU.
Example	recipes/skillclaw/

What it does #

A day is the frozen 60 task list run once against the pool the night before left behind. A night reads that day's recorded traffic and makes one decision per skill group plus the no-skill bucket: create, improve, merge, or skip. Every non-skip decision applies, so the pool the next day measures is always the one the night produced.

The method needs no separate judge of its own edits. The next day is the measurement, which is why the control run matters: it replays the same rounds and reports nothing, so its pool never changes and its scores are the baseline the method run is read against.

How Reef implements it #

skillclaw is a method package on the harness evolution engine (reef/train/cordis_backend/); Evolve your harness describes the mechanism it runs on. The method supplies a Proposer and an EpisodeScorer (harness/plugins.py); the engine renders the harness, runs evaluation tasks, records changes, and rolls back rejected changes. propose runs the night flow and maps its decisions to one composite mutation sequence, so a whole night applies under one snapshot and settles under one verdict.

Two choices distinguish it from the tutorial method. selection: always publishes every night that produces a proposal, because the next day is the gate. And the skill catalog must reach the model on every live request, so the recipe overrides build_surface, the hook controlling how a published artifact reaches the request path (recipe.py). That override is the one extension point beyond the method's own plugin classes; with it the recipe names no harness_adapter, since no harness tree is installed.

Configuration #

recipes/skillclaw/skillclaw.yaml is the recipe config the driver boots. It names recipes.skillclaw.recipe:SkillClawRecipe as its implementation and sets pipeline.task2batch.config.batch_size: 60, so every report of the day batches and the day's last one triggers the night. The engine's keys are in Recipe configuration.

Run the example #

recipes/skillclaw has the prerequisites and the walkthrough. REEF_SC_RUN selects the run: skillclaw is the method run, frozen the control.

Results #

The gain criterion is fixed before any data is read: a category's gain counts as real only when the method run's final day beats the control mean by more than two control standard deviations, one sided, calibrated to a 5.6 percent false positive rate per category.

On GLM-5.3-Flash served locally on four GPUs, six nights applied 13 skill improvements and 8 creations, and the pool grew from 9 to 17 skills. In the Productivity category the method run's final day beats the control mean by +12.05 points (2.29 sd); on the subset of tasks scored on every day in both runs the margin grows to +15.72 (2.50 sd). python run.py report recomputes the numbers from the stored run data.