Working draft

PRO-05 Agent Incident, Suspension and Retirement Procedure

Stop first, understand second: three switches to stop an agent, severity rules, incident steps, and how to retire an agent completely.

Working draft, published openly. This is the version Temei works to. You may read it in full, adapt it and use it in your own organisation. If you would like help fitting it to your tenant, book a call.

FieldValue
OwnerFounder, FIDEVO/AI LABS LIMITED (trading as Temei)
StatusWorking draft, published for open use
Version0.1 draft, 22 November 2026
Reviewyearly, and after every Severity 1 or 2 incident
Applies toevery Temei-built agent and its workflows, in Temei's tenant or a client's
Sits underPOL-01 Agent Autonomy Policy (section 6, rule A8) · POL-02 Agent Data and Privacy Policy (rule D8)

When an agent does something wrong, stop it first and understand it second. When an agent is no longer needed, remove it completely and keep the record. This procedure covers both.

1. How to stop an agent

Agents built to STD-01 can only write through their write workflow and only run when a trigger workflow starts them. That gives three switches, from fastest to widest:

SwitchWhat it stopsWhereWho
S1 Turn off the write workflowevery write the agent could make; the agent may still run but changes nothingCopilot Studio → Workflows → the agent's write workflow → turn offowner or platform admin
S2 Turn off the trigger workflow(s)every automatic runCopilot Studio → Workflows → the trigger workflow → turn offowner or platform admin
S3 Quarantine or block the agentall use of the agent, including chat and channelsCopilot Studio quarantine (Power Platform API or pac copilot-studio set-bot-as-quarantined); Microsoft 365 admin centre → Agents → Block, for agents in Copilot or Teamsplatform admin

For any suspected harm, use S1 and S2 together. They take minutes and are fully reversible. Use S3 when the agent is published to people or the cause is unknown. The register entry names every workflow, so nobody has to search for them under pressure.

2. Severity

SeverityExamplesStop withinTell
1 Criticalpersonal or confidential data exposed; wrong live writes at scale; agent acting outside its lists; anything that may need reporting to the ICO15 minutesowner and approver at once; client's accountable owner and DPO at once
2 Majora scope breach without harm; one wrong write on live data; duplicate check failed1 hourowner and approver the same day; client contact the same day
3 Minorfailed runs; odd or wrong output that changed nothing live; cost spikenext working dayowner; noted in the weekly review

When unsure, choose the higher severity. It can be lowered once the facts are known.

3. Incident steps

  1. Stop. Apply the switches from section 1 for the severity. Note the time.
  2. Record. Open an escalation titled Incident: <agent> <short description> with severity, time found, who found it and what was switched off.
  3. Contain the data. List the items the agent wrote or changed during the window (query by author and time). Correct or remove wrong items, keeping a copy for the review.
  4. Assess personal data. If personal data may have been exposed, the controller decides within 72 hours of becoming aware whether to report to the ICO, and whether to tell the people affected. Temei, as processor, supplies the facts without delay (POL-02 D8).
  5. Demote. Apply POL-01 section 6: scope breach means back to L1; harm or exposure means suspended. Update AutonomyLevel and AgentStatus in the register.
  6. Find the cause. Instructions, tool settings, workflow logic, data change, platform change or misuse. Check the run history of each workflow.
  7. Fix in the sandbox. Change, then re-run STD-01 tests T1 to T6, plus a new test that reproduces the incident.
  8. Review. Within 5 working days for Severity 1 and 2: what happened, why, what changed, what would have caught it earlier. Write it in the escalation and in Logos.
  9. Reinstate. A suspended agent returns only through PRO-02, starting from the level POL-01 allows. A Severity 3 fix may return to the same level after the owner checks T1 to T5.
  10. Close. Resolve the escalation with the fix in Resolution. The Knowledge Agent will draft an article from it; a person approves it before it is published.

4. Suspension without an incident

Suspend the agent with S1 and S2 (no review needed) when:

  • its owner leaves and no new owner is named within 5 working days;
  • its review date is more than 30 days past (POL-01 also demotes it one level);
  • its data policy changes and the needed connectors are no longer allowed;
  • the client asks, for any reason.

Record the suspension in the register and in Logos. It returns when the cause is fixed and the owner has run T1 to T5.

5. Retirement

Retire an agent when its purpose has gone, it has been replaced, it has been suspended for more than 60 days, or it costs more than it saves.

  1. Decide. The owner proposes, the approver agrees; write it in Logos with the reason.
  2. Tell the people who use it or read its output, with the date it stops.
  3. Stop. S1 and S2, then wait one full cycle to make sure nothing still depends on it.
  4. Unpublish. Remove channels and sharing; block it for the tenant if it was in Copilot or Teams.
  5. Delete the workflows, write workflow and trigger workflows.
  6. Delete the agent in Copilot Studio. Where the agent has a Microsoft Entra agent identity, deleting it removes that identity too; confirm it is gone.
  7. Remove connections created only for this agent, and any access granted only to it.
  8. Deal with its data under POL-02 section 5: the items it wrote stay under the site's retention unless the controller decides otherwise; nothing is kept elsewhere.
  9. Update the register: AgentStatus Retired, retirement date, and a link to the Logos entry. The register item is never deleted.
  10. Archive the record. The gate, Registration Card, certificates and incident reviews are kept for 6 years after retirement.

6. Post-incident review (template)

FieldEntry
Agent and level
Severity1 / 2 / 3
Founddate, time, by whom, how
Stoppeddate, time, switches used
What happenedfacts, in order
Impactitems affected; personal data yes/no; people affected
ReportedICO yes/no and date (controller's decision); people told yes/no
Cause
Fixwhat changed, and the new test that covers it
Earlier warningwhat would have caught it sooner
Level after
Reviewername, date

7. Practice

Once a quarter, for one agent at L3 or above, the owner runs a stop drill: switch S1 and S2, time it, switch back and confirm the agent works. Record the time in the register. For agents at L1, the drill is part of STD-01 test T6.

9. Approval

When you adopt this document, record who approved it, their role and the date, and keep that record with your other governance records.