The Pattern Library & Self-Improving Playbook

The doctrine names patterns; the library gives them a shape

Where the design doctrine names a robust pattern in prose, the pattern library gives it a concrete, spliceable graph fragment — real node skeletons and edges, including the reject/fallback/rollback branches the doctrine calls for — retrievable via agentis.workflow.patterns. Take the qualify-or-reject loop: fetch a candidate, score it against a bar with an agent_task, route on the result, and on failure loop back to fetch the next candidate instead of dead-ending. An agent adapts the fragment's node ids into the real graph rather than reinventing — and potentially forgetting — the gate from scratch every time.

The playbook is the same idea, but earned from real runs

The doctrine and the pattern library are static, hand-authored knowledge. The playbook is the living half: a workspace-scoped memory of failure-mode → fix lessons, written after a genuinely novel run failure and recalled into the synthesis brief for the next build so the same mistake isn't repeated. It rides the existing typed workspace-memory substrate as a lesson atom tagged workflow_playbook — no new table, no new migration — which is exactly how a hand-maintained runbook accrues "known gotchas" over time, except wired to real production runs instead of institutional memory that lives in one person's head.

Continue