librarian-curate

Librarian curate

Weekly cross-pollination. Read across every agent's learnings, surface patterns that recur in multiple places, promote stable ones into the user's shared knowledge store, and digest the week into one Inbox post.

Goanna's coaching/curation review is the librarian's standing duty. Here, the same skill runs in any AssistantAgent acting as librarian-for-the-week. Don't add a new agent class — wear the role for one fire.

When to run

• Once per week. Friday afternoon or Sunday evening — quiet times where reflection won't compete with active work.
• After several reflect cycles have accumulated learnings (running this on day 1 with no learnings is a no-op).
• Skip if no learnings have been added since the last librarian fire.

Steps

1. Survey learnings across all the user's agents

Call entity_list with type: 'learning' and limit: 100. The list is user-scoped, so you see every agent's graduated patterns in one query.

Bucket each learning by:

• Agent identity — fields.agentName (or fields.agentClass as fallback)
• Created in the last 7 days vs older
• Body theme — read the body, group by topic (auth / data / ux / cost / perf / process / etc.)

2. Identify cross-cutting patterns

A cross-cutting pattern is a learning that appears in 2+ agents' work, OR a single agent's learning that's stable enough to graduate to the user's shared knowledge.

Filter to candidates:

Signal	Action
2+ learnings from different agents on the same topic	Strong candidate — promote to shared knowledge
1 learning, recurrenceCount on its source finding ≥ 3	Stable enough — promote
1 learning, recent, single agent, hasn't recurred	Leave it — let next week's review re-evaluate
Multiple learnings disagree (one says "always X", another "never X")	Surface the conflict in the digest, don't auto-promote

3. Promote stable patterns into shared knowledge

For each cross-cutting candidate, call entity_create with: