AWS has published a worked example of turning an AI incident investigation into a proposed infrastructure fix, with human approval required before changes are applied. The key design choice is a clear boundary: read-only actions can run automatically, but changes to production infrastructure pause for an engineer.
AWS AI Watch analysis
What happened
The example starts with AWS DevOps Agent investigating an incident and sending its findings to an EventBridge-triggered workflow. A durable Lambda function passes the investigation to Amazon Bedrock, which selects from a curated allowlist of purpose-built Lambda tools to propose a remediation.
AWS’s walkthrough demonstrates the process using a test Lambda function whose three-second timeout causes failures. The proposed remedy is to raise that timeout to 30 seconds. Read-only checks can run without approval; a change to infrastructure suspends the workflow until a person approves or rejects it.
Why it matters
This is a concrete recipe for extending an AI agent from diagnosing an incident to preparing a fix, without giving the model unrestricted permission to alter production systems. The allowlist constrains which tools it can call, while the approval gate separates investigation from consequential action.
The durable function can checkpoint and wait for an approval signal, then resume. That makes the hand-off part of the workflow rather than a race against a short-lived process. AWS presents this as an example built from its services, not evidence that the approach is risk-free or suitable for every production environment.
Our read
This is the more persuasive version of “AI agents will fix everything”: let the system gather evidence and prepare a bounded change, then make a person own the go/no-go. The useful test for operators is not whether the model sounds confident, but whether the tools are narrow, the proposed change is reviewable and the approval step is real.
What to watch
- Whether AWS or customers share results from deployments beyond the walkthrough’s test incident.
- How teams review proposed changes and handle rejections or parameter overrides.
- Whether the allowlist and approval boundary remain clear as workflows grow more complex.
Discussion spark: For production incidents, should AI be allowed to apply tightly scoped fixes automatically, or should every change wait for an engineer’s approval?
Sources and evidence
- Automate remediation post AWS DevOps Agent investigation (7 October 2026, 15:46 UTC)
- How Cornerstone OnDemand cut database diagnosis by 78% with Amazon Bedrock (7 October 2026, 15:46 UTC)
- Building AI builders: Playbook for closing the AI knowledge-capability gap (7 October 2026, 15:46 UTC)
not affiliated with or endorsed by Amazon Web Services (AWS)