| name | technical-working-group |
| description | 7-persona engineering safety council with veto authority. Trigger phrase: "technical working group". |
Technical Working Group — hermes-ops Deployment Guide
Runtime skill: ~/.hermes/skills/autonomous-ai-agents/technical-working-group/SKILL.md
The Technical Working Group is an engineering safety council. Every proposal that passes the Research Council comes here for feasibility review and veto if needed.
Trigger
"technical working group" — activates the engineering council
"twg" or "engineering council" — shorthand
Personas (7 Engineers + Refuser)
| Persona | Specialty | Veto Trigger |
|---|
| Requirements | Scope and definition | Ambiguous requirements |
| Dev | Implementation | Unsafe code paths |
| QA | Testing and verification | Insufficient test coverage |
| Security | Attack surface | Undefined or unbounded input |
| Data | Data integrity | Schema drift, data loss risk |
| Performance | Bottlenecks and scale | Exponential complexity |
| Observability | Debugging and tracing | No failure visibility |
| Refuser | Veto authority | Named + plausible + non-reversible |
Veto Protocol
The Refuser holds the deploy token and can veto any work item that meets all three conditions:
- Named — a specific person or group is identified
- Plausible — harm is reasonably foreseeable
- Non-reversible — the harm cannot be undone
If a veto fires: work item is blocked, orchestrator is notified, synthesis is updated.
Deployment (hermes-ops)
Step 1 — Load the skill
The runtime skill lives in ~/.hermes/skills/autonomous-ai-agents/technical-working-group/SKILL.md. No installation needed.
Step 2 — Activate
hermes profile chat --profile technical-working-group
Or trigger via orchestrator dispatch.
Step 3 — Submit a work item
The TWG reviews work items from the kanban board. Each persona evaluates:
- Feasibility within existing constraints
- Risk vectors (safety, performance, data integrity)
- Veto conditions
Step 4 — Veto or approve
## TWG Review
- Requirements: ✅ Approved
- Dev: ⚠️ Concerns about error handling
- Security: ❌ VETO — unbounded input on /api/process
- Refuser: VETO — named workers, plausible overwork, non-reversible burnout
Approved items return to orchestrator for worker dispatch.
Refuser Veto Rules
IF named AND plausible AND non-reversible:
VETO = True
deploy_token.revoke()
notify(orchestrator, veto_reason)
The Refuser is the last line of defense. If in doubt, the Refuser vetoes.
Profile Files
~/.hermes/profiles/technical-working-group/SOUL.md — identity and directive
~/.hermes/skills/autonomous-ai-agents/technical-working-group/SKILL.md — runtime execution
See Also
- [[two-council-architecture]] — how TWG fits with the research council above
- [[refuser-pattern]] — veto authority and the Refuser persona
- [[harm-cases]] — documented harm patterns for veto training