Agents Catalog
Agents are heavy-lift task handlers that run in isolation from the main conversation. Complex or resource-intensive work — parsing large files, searching the knowledge graph, assembling session summaries — happens in a separate context so it does not crowd out working memory. Skills and commands trigger agents automatically when the work warrants it; direct invocation is rarely needed.
Agent Reference
| Agent | What it does | Invocation |
|---|---|---|
| lesson-capture-agent | Real-time lesson capture from active sessions | Auto-triggered by lesson-capture skill |
| session-summary-agent | Session summaries with open plans and ADRs tracked | Auto-triggered by session-wrap skill or /kmgraph:session-summary |
| recall-agent | Natural-language search across the knowledge graph | Auto-triggered by kg-recall skill or /kmgraph:recall |
| rules-capture-agent | Dedup check against target file, draft new rule in house style (Always/Never + Why/Source), approve/edit/discard loop, write to one of 4 targets, MEMORY.md pointer stub | Auto-triggered by rules-capture skill (implicit corrections) and capture-router skill (explicit behavioral corrections) |
| knowledge-extractor | Large-file parsing for KG extraction (approval-gated writes) | --delegate knowledge-extractor flag or auto for large operations |
| knowledge-reviewer | Quality review for lessons and ADRs before saving | Auto-triggered on capture in review mode |
| session-documenter | Git archaeology for complex multi-branch sessions (approval-gated commits/pushes) | --delegate session-documenter flag |
| platform-sync-agent | Cross-platform config file management | Triggered when platform config files change |
| mcp-setup-agent | IDE detection and MCP server registration | /kmgraph:setup-platform |
Approval-Gated Agents
The following agents never write files or push commits without user approval:
- knowledge-extractor — Read-only parsing. Presents extracted content for review before writing to the knowledge graph.
- session-documenter — Assembles session summaries. Never auto-commits or auto-pushes.
Invoking Agents Explicitly
Most agents are invoked automatically. To delegate a heavy operation explicitly:
/kmgraph:update-graph --delegate knowledge-extractor
/kmgraph:session-summary --delegate session-documenter
/kmgraph:extract-chat --delegate knowledge-extractor
Related
- Skills Catalog — Auto-triggered context providers that invoke agents
- Command Guide — Full command reference
- Concepts — Four-Layer Architecture — How agents fit the architecture