What we investigate
An agent incident is often an identity and authorisation incident as well as a model-behaviour problem. Establish the agent version, execution context, delegated identities and connected tools.
Response work to agree
- Preserve run IDs, tool arguments, responses, approvals and workflow versions.
- Determine which actions were attempted, denied, completed or retried. Check the destination system rather than relying only on generated text.
- Agree isolation of the affected workflow, scheduled jobs, connector or credential with its owner.
- Inventory queued work and side effects before reconnecting the agent.
What safe recovery requires
A stopped agent may leave valid tokens, queued jobs or changed records behind. Recovery includes reviewing those residual effects and restoring only the permissions the task needs.
Start with a non-sensitive brief
Use the incident brief builder or contact the team. Describe the affected service, observed impact and whether activity is ongoing. Availability, authority and delivery scope are confirmed before response begins.
Turn observations into investigation questions
| Observation | Evidence to seek | Decision to resolve |
|---|---|---|
| An agent edited a record outside its intended task | Run, delegated identity, record history and approval | Which identity completed the change? |
| A task continued after the agent was paused | Scheduler, queue, retry and external job records | Which side effects remained active? |
| Several agents used the same credentials | Service principal, credential ownership and tool access | Can the affected workload be isolated without stopping the others? |
Account for work already launched
List external jobs, scheduled runs and retries created before the pause. Confirm the status of each through its owner. Reversing a completed financial or customer-facing action may require a business process rather than a technical rollback. Record who owns each residual effect before restart.
What we agree to deliver
The scope identifies the questions to answer, records available, customer owners and permitted changes. Typical outputs include an event timeline, the affected boundary, approved containment actions, recovery checks and unresolved risks. The deliverables page explains how findings connect to evidence.
Continue with agent forensic evidence, or compare response engagement options.
