Learn
What is the agent lifecycle?
The agent lifecycle covers design → eval → deploy → operate → improve → retire for each role — like software, not a one-time prompt. Skills version in git; promotion requires passing evals and named ownership.
In depth
Treat roles like services:
- Design — queue, connectors, prohibited actions, owner, KPI - Build — skill + eval set in git - Staging — run eval pack; red-team injection cases - Production — heartbeats enabled; Ask on writes - Operate — weekly metrics, sample review, budget check - Improve — skill PRs from correction categories - Retire — revoke connectors, archive skill, export audit history
Skipping lifecycle steps is how teams get 'works in demo, fails in week two.' Promotion gates — eval pass rate, owner sign-off — prevent untested skill changes from reaching customer-facing roles.
Neuro OS aligns lifecycle with engineering practice: versioned artifacts, receipts, and rollback.
Examples
- New support role — 30 ticket eval before production toggle
- Skill v1.3 — PR review + eval CI green + owner approve
- Role retired — CRM connector revoked; logs archived
- Pilot → prod — Ask required in pilot; selective pre-approve in prod
Related terms
FAQ
How often update skills?
When correction categories repeat — not on every model vendor release.
Who approves promotion?
Named role owner + optional security for new connectors.
Run governed agent roles on a company OS — not only definitions in a glossary.