| name | agent-learning |
| description | Use when agent-side working philosophy, interaction lessons, task retrospectives, repeated routing misses, missed skill invocation, or recurrence-prevention feedback should be logged without mixing them into user preferences. |
Agent Learning
Tool Commands
Use the command packet before applying this skill's workflow:
python3 tools/agent_tools/skill_tool_commands.py show --skill agent-learning --format text
Execute the required and task-matching conditional commands that the packet prints.
- Read
agents/skills/agent-learning.md.
- Read
agents/workflows/agent-learning-workflow.md.
- Treat plain
agent-learning or $agent-learning in a user request as an explicit skill invocation, not only a candidate signal.
- Select this skill for user / reviewer feedback about agent behavior, repeated misses, missed skill invocation, over-constrained routing, recurrence prevention, "こういう止まり方", task retrospectives, or agent-side memory updates.
- Separate user preference from agent-side learning.
- Record observable behavior with
python3 tools/agent_tools/workflow_monitor.py --report-dir reports/agents/<run-id> --behavior-event "...", including skill invocation, subagent routing, tool gates, prompt evals, review feedback, subagent lifecycle, and diff-check decisions.
- Record user or reviewer feedback observed during actual use with
python3 tools/agent_tools/workflow_monitor.py --report-dir reports/agents/<run-id> --runtime-feedback "source=<user|reviewer|eval> target=<skill-or-workflow-or-eval> action=<prompt_repair|eval_update|memory_record|no_op>", then update the targeted skill prompt, workflow prompt, eval, or memory note before closeout.
- When feedback says the currently used skill is weak, shallow, late, or misrouted, treat the active skill as the first repair candidate and confirm the owner before editing. Before turning feedback into a durable rule, calibrate the strength of the change; prefer scoped guidance or examples when one observation does not justify a hard rule. Reserve hard rules for invariant, checker-backed, or repeatedly observed failures. If the skill prompt changes, update both the discoverable runtime
SKILL.md and the canonical agents/skills/<skill>.md owner, then update or verify the matching prompt eval.
- Treat user or reviewer feedback about passing tests by simplification, revert, intended-behavior deletion, oracle weakening, or over-weighting test planning until owning code repair stalls as active skill feedback for
test-design and the implementation workflow. Record it with workflow_monitor.py --runtime-feedback targeting test-design or the implementation workflow and action=prompt_repair|eval_update, then resolve it through prompt, eval, or tool repair. Do not close it as memory-only.
- Treat user or reviewer feedback that algorithm repair started from test
edits, expected-value changes, tolerance changes, or oracle changes before
the algorithm contract and code-side repair route were fixed as active skill
feedback for
computational-optimization, algorithm-proof-exploration,
test-design, and the implementation workflow. Resolve it by repairing the
prompts so algorithm contract, public entrypoint, recurrence or state
transition, invariant, stopping or acceptance rule, failure semantics, and
selected code repair route come before test changes.
- Treat feedback about excessive input tokens, repeated context loading, duplicated raw logs, or broad prose being passed into the model as active routing/context-skill feedback first. Repair the selected skill by preserving owner/dependency evidence while routing structure reading through the protocol-owned
Structure Intake Packet and moving duplicated or bulky raw material to artifacts plus structured summaries; do not convert the feedback into a blanket rule to omit needed context.
- Before closeout, run
python3 tools/agent_tools/evaluate_agent_run.py --report-dir reports/agents/<run-id> --behavior-manifest evidence/agent-evals/agent_behavior_eval.toml --write and resolve any feedback actions.
- Use
$result-artifact-writeout when eval, hook, or skill feedback results need durable writeout; keep raw evidence, summary, manifest, and unique artifact path separate.
- Log concise evidence-backed observations with
tools/agent_tools/log_agent_learning.py.
- If memory files changed, run
python3 tools/agent_tools/persist_agent_memory.py --commit --push and, from a template superproject, also close the vendor/agent-canon pin with --commit-superproject --push-superproject or an equivalent explicit commit.
- Keep raw chat out of notes; record source, evidence, scope, and confidence.
- Promote only stable items into
AGENTS.md, workflow docs, or guardrails.