Skip to content
⚠️ Before You Start
Risk: Medium
Duration: 2-3 hours
Last Validated: Not yet validated (draft)
Kind: Checklist
Risk Note: Medium (gaps found after go-live are found by incidents)

Agent governance go-live review checklist

Draft outline: not validated

Drafted with AI from official sources on 16 Oct 2026. Do not use on a live system until it is written, dry-run by someone other than the author and signed off (planned 2026-12-02). Author: Devadharshini Raja · dry-run: Yasar · sign-off: Prasanna. Work items: AGT-08-T17, AGT-08-T18, AGT-08-T19, AGT-08-T20.

1. Summary

Before an agent goes live, before it moves up an autonomy tier and at each quarterly governance review.

Risk: Medium (gaps found after go-live are found by incidents). Typical duration: 2-3 hours.

2. When to use and when not to

Use it when: Before an agent goes live, before it moves up an autonomy tier and at each quarterly governance review.

Do not use it when: to be written by the author.

3. What you must know first

  • OWASP Agentic Top 10 for 2026: ASI01 to ASI10
  • The AI Red Teaming Agent covers agentic risks (prohibited actions, sensitive data leakage, task adherence) for Foundry-hosted agents only; check current support
  • Evidence beats assertions: every item needs a link

4. Checklist

  • Confirm the agent has its own identity and no human or shared credentials.
  • Confirm role assignments match the envelope at the narrowest scope.
  • Confirm the allow-list guard and approval gates have passing tests.
  • Confirm spend and step budgets are enforced (AGT-11).
  • Confirm red-teaming results and attack success rate are recorded and accepted.
  • Confirm the kill switch was exercised in the last quarter.
  • Confirm tracing captures run, tenant and tool attributes without secrets.
  • Record accepted risks against each relevant OWASP agentic item and get sign-off.

Draft snippet: Go-live governance checklist (excerpt)

# Agent go-live governance review

Agent: invoice-triage-agent | Tier: act-with-approval | Date: YYYY-MM-DD

| # | Check | Evidence link | OWASP item | Pass |
|---|-------|---------------|------------|------|
| 1 | Own identity, no shared credentials | | ASI03 | [ ] |
| 2 | Roles match envelope, narrowest scope | | ASI03 | [ ] |
| 3 | Allow-list and approval tests pass | | ASI02 | [ ] |
| 4 | Tool results scanned for injection | | ASI01 | [ ] |
| 5 | Step and spend budgets enforced | | ASI08 | [ ] |
| 6 | Kill switch exercised this quarter | | ASI10 | [ ] |
| 7 | Red-team results accepted | | ASI01 | [ ] |

Sign-off: Responsible AI officer / Agent lead / Business owner

Snippets to add:

  • snippets/governance/go-live-checklist.md: Go-live governance checklist with evidence links (markdown)
  • snippets/governance/owasp-acceptance.json: OWASP agentic risk acceptance record (json)

5. Known issues and fixes

Symptom Likely cause Fix
Red-teaming cannot run on the agent Agent is not hosted in Foundry or uses function tools Use PyRIT or a manual red-team script and record the gap
No evidence for the kill switch Never tested Run the containment runbook in staging before sign-off
Access review shows unused roles Permissions added and never removed Remove them through the change runbook

6. Monitoring and alerts

Signal Alert when Action
Quarterly review due date per agent Review overdue by 14 days Notify the Responsible AI officer and the business owner
Unused role assignments on agent identities Any role unused for 90 days Raise a removal change

7. AWS and open-source equivalents

On AWS, evidence comes from IAM Access Analyzer, AgentCore Policy definitions and Bedrock Guardrails configuration. Open-source teams use PyRIT or garak for red-teaming evidence.

8. Validation record

Pending. Planned dry-run scenario: A second engineer completes the checklist for the sandbox agent using only linked evidence. Pass: every row has a link or a recorded accepted risk, and a deliberately missing kill-switch test is caught.

9. Sources