self-improving-agent

Turn hard-won lessons into tested behavior changes.

agent improvement

What it does

Turns evidence from completed work into bounded, auditable behavior changes that are captured, validated, applied, or rejected.

When to use it

Use after recurring failures or workarounds, user corrections, validated successes, or requests to review learning candidates.

Give it a completed-work lesson or candidate-review request; every run records one explicit lifecycle outcome and may install a validated change.

This skill

agent skill files

Write

Edits one owning skill

pnpm dlx @codeharbor/agent-playbook

Execute

Installs skills and failure hooks

Claude Code failure hook

Write

Enables the Claude failure hook

Claude Code failure-hook events

Read

Captures failed tool events

The workflow uses the apb CLI to capture, inspect, evaluate, review, apply, and export lessons.