What happened
The New Stack published a piece on AI SRE agent autonomy, arguing that most AI incident tools still leave a key decision to humans. The article frames incident response in three phases and asks whether you can trust an agent to decide on its own, not just help with troubleshooting. The text we could read was partial, so we are not repeating its detail here. Read the full piece for its specifics.
Why it matters
For a company without a platform team, an AI agent that reads logs and suggests a cause is easy to try. The harder question is what happens when it is allowed to act: restart a service, roll back a release, fail over a database. Those actions can fix an outage or make it worse.
The risk sits with whoever is on call. That is often one or two engineers, or a founder who still has production access. If an agent acts on a wrong guess at 3am, nobody may be awake to notice. If nobody has written down which actions are safe, the agent’s autonomy gets decided by accident rather than by choice.
Recovery is the part that gets tested hardest. An agent can only safely take an action if you know it can be undone, and that depends on backups, rollbacks and failover that work in practice, not only on paper.
What to do this week
- Write down the actions your on-call person takes most often during incidents, such as restarting, rolling back, scaling or failing over.
- Sort them into three groups: safe to automate, needs a human to approve, and never automated. Be strict about anything that deletes or overwrites data.
- For each action you would let an agent touch, check that it can be reversed, and that you have actually tested the reversal.
- If you are trialling an AI incident tool, start it in read-only mode. Let it suggest, and compare its suggestions with what your engineers did.
- Check that the agent’s credentials are narrow and that every action it takes is logged where you can review it.
Need a hand?
Automation in incident response is only as safe as the recovery plan underneath it. Our Disaster Recovery work covers backups, restore tests, failover and the runbooks that say who decides what. The first call is free and it is with an engineer, not a salesperson.
Source: Most AI incident tools still leave a key decision to humans
Related service: Disaster Recovery & Business Continuity