Three background agents keep the knowledge base healthy without requiring constant human curation. Each has a distinct trigger, a distinct risk profile, and a distinct write mode.
Three agents, three write paths. Risk determines whether the write is proposed (human-in-loop) or direct (revision-undo).
Capture agent
TRIGGER
Fires automatically after every agent response. Runs in the background so the user never waits.
CONTEXT
Receives the full conversation (up to 15 turns), the agent synthesis, and a summary of existing knowledge base pages in scope.
DECISION
Three possible outcomes: CREATE a new page, UPDATE an existing page, or NOTHING if the synthesis is ephemeral or already covered.
WRITES
Creates a proposal with status pending. Users review, edit, accept, or dismiss proposals in the UI. Accepted proposals carry full source attribution.
MODEL
A cost-optimized model designed for speed -- one call per conversation. Fast enough to run on every conversation without budget concern.
Health agent
TRIGGER
Automated nightly schedule. The structural health scan itself requires no AI calls. When a scope accumulates 5+ issues, a maintenance task dispatches the health agent to fix them.
RED-LINK HANDLING
Finds or creates pages that satisfy broken [[references]], restoring graph connectivity. New pages are auto-linked.
ORPHAN HANDLING
Identifies pages with zero inbound links and proposes connections to related content, integrating isolated pages into the graph.
THIN CONTENT
Flags pages with low citation counts or minimal source references. Under-sourced pages lack the attribution needed for trust.
STALENESS
Detects pages not updated beyond a configurable threshold. Checks whether cited sources have newer versions and flags sections for re-evaluation.
WRITES
Direct writes with full revision history. Every change is reversible via the revision history. Structural fixes (link resolution, orphan connection) are low-risk and deterministic.
Compiler agent
The compiler agent is the primary ingestion path for external documents. It reads connector items, chunks and embeds them, merges content into knowledge base pages with blame and citation metadata, and creates revisions. For a detailed walkthrough of the compilation pipeline, see the knowledge base compiler page.
Human vs agent updates
Four write paths, one revision history. Anything can be rolled back via the revision history.
Human review for risky content changes.
Higher-risk content changes stay in a human review queue before they become part of the knowledge base. The capture agent uses proposals because it creates or substantially modifies content. The health agent writes directly because its changes are structural (fixing links, connecting orphans) and fully reversible through the revision history.
