Autonomous Curation Agents (epic tracker)

[rdwj]

Summary

Epic tracker for the autonomous curation agent fleet: four specialized agents (Curator, Fact Checker, Trace Reviewer, Statistician) running on-cluster as Kubernetes Deployments/CronJobs, each with scoped RBAC, independent model selection, and independent enable/disable.

Design reference

• planning/autonomous-curation-agents.md
• docs/curator-agent.md (Phase 3 background curator sketch)
• docs/knowledge-compilation.md (Design: Knowledge compilation — governed LLM-maintained knowledge synthesis #171)
• docs/conversation-persistence.md (Design: Conversation thread persistence as governed MemoryHub subsystem #168)
• docs/governance.md (service agent identity model)

Phases

• Phase 0: Prerequisites (relevant_until schema, Valkey deployment, OBO authorization)
• Phase 1: Curator Agent (deep dedup, staleness, conflict detection)
• Phase 2: Fact Checker Agent (temporal expiry, verification plugins)
• Phase 3: Trace Reviewer Agent (post-session extraction from threads)
• Phase 4: Statistician Agent (population patterns, convergence signals, SDC)
• Phase 5: Promotion pipeline integration (five-stage pipeline)
• Phase 6: Fine-tuning and optimization

Cross-cutting

• Shared agent framework (authenticate/dequeue/process/report)
• Within-user pattern surfacing (search-time pattern signals)

Dependencies

• Design: Conversation thread persistence as governed MemoryHub subsystem #168 (conversation persistence) -- required for Trace Reviewer full mode
• Design: Knowledge compilation — governed LLM-maintained knowledge synthesis #171 (knowledge compilation) -- complementary, shares Valkey
• Governance service agent identity model -- implemented
• Inline curation pipeline (Phase 2a) -- implemented