The agent writes the skill it wished it had

Three nights in a row the overnight agent hits the same problem: a Terraform state lock that needs a specific unlock-and-retry sequence. Three nights in a row it rediscovers the sequence from scratch.

Its prompt carries a standing instruction: when you solve the same class of problem a third time, write it down as a skill. The agent POSTs /api/v1/agent/skills to create tf-state-unlock, then PUT /api/v1/agent/skills/source with the steps, the checks, and the failure modes it learned.

The new skill does not run itself. It lands in the registry with a trust state you control. Over coffee you read the source and POST trust and toggle. On night five the agent reaches for the skill instead of the search.

Runbooks used to be written by whoever got paged. The agent writes its own playbook; you decide which pages count.

The agent writes its own playbook; you decide which pages count.