Handbook
Workcell model
An agent workcell is a bounded agent runtime that accepts a WorkcellRequest, performs scoped work under an approval mode, and returns a WorkcellResult linked to a ForgeRun.
Definition
An agent workcell is a bounded agent runtime that accepts a WorkcellRequest, performs scoped work under an approval mode, and returns a WorkcellResult linked to a ForgeRun.
Workcell types
operator— orchestration, planning, coordination.implementation_worker— coding, patching, review.reviewer— second-pass check.channel_gateway— chat/mobile ingress and relay.versona_runner— discipline-lens review.evidence_summarizer— converts outputs into human review packets.
Current workcells
| Workcell | Best role |
|---|---|
| Hermes | adaptive operator with memory, skills, and MCP orchestration |
| Factory Droid | implementation and code review worker |
| OpenClaw | messaging/channel ingress and status relay |
local_runner |
Self-Host Alpha bounded repo checks without Fleet |
campaign_orchestrator |
Forge Campaign lifecycle on the operator host or CI (branch, Fleet API, git, PR stack) |
fleet_ux_worker |
UX audit/remediation inside a Fleet container (Playwright, optional headless agent) |
cursor_cli |
Cursor CLI agent steps when governed under a ForgeRun |
local_llm_worker |
bounded local-LLM judgment (micro-pack + LCDL); runner in forge-workcells (private repo) |
| Future agents | plug in through WorkcellRequest/WorkcellResult |
Runner homes: local_runner remains in forge-platform (scripts/selfhost_runner.py) for Self-Host Alpha. local_llm_worker lives in forge-workcells; see Forge micro-agent.
See Forge Agent for definitions and boundaries.
Rule
The platform should still work if Hermes is disabled. Hermes increases convenience and learning; it must not be required for governance.