Cyborg operating model

What's on this page

• Human supervision model
• Audit and retention expectations
• Kill switches and break-glass
• Evaluation before production enablement

What this enables

• Safe scale-out of agents across enterprise functions

Supervision

Tier	Supervision
CRITICAL	Human approval every side-effecting action
HIGH	Approval for prod/security/legal; async review otherwise
MEDIUM	Sampling review + automated policy checks
LOW	Automated only with rate limits

Audit

• Log: actor (human), cyborg id, tools invoked, decision summary, timestamp
• Retain per document retention policy and legal holds
• FO logs segregated and access-controlled

Kill switches

1. Disable namespace admission (deploy layer)
2. Revoke tool credentials
3. Mark cyborg priority drained / offline in catalog
4. Incident response if misuse suspected

Evaluation checklist

• [ ] Prompt matches job responsibilities
• [ ] Tools are least privilege
• [ ] knowledge_docs are current
• [ ] Security level correct
• [ ] Red-team / abuse cases considered for GTM and support bots