Reliability

Agents Need Guardrails, Not Just Goals

It’s been a busy few months for agentic AI announcements, at my employer and across the industry:

  • In February Kyndryl introduced a policy-as-code capability for agentic AI workflows. Rules and regulatory requirements are turned into machine-readable policy, so agents can only take approved, predictable actions.
  • On 10 June Kyndryl launched AI Orchestration for Business, built on the Kyndryl Agentic AI Framework. It coordinates agents across systems, with guardrails and human oversight built in.
  • On 18 June Kyndryl expanded its collaboration with AWS on enterprise agentic AI adoption, including being a launch partner for the AWS European Sovereign Cloud.

What Changes When the Operator Is an Agent?

Less than you might think. And that’s the point.

Everything SRE teaches about safe change applies to agents. Arguably more so, because agents work faster, never get tired, and do exactly what they’re allowed to do, not necessarily what you meant. Back in February I mentioned reports of an AI coding agent involved in a production interruption. The fix there wasn’t “less AI”. It was better access control and review.

My checklist for agents in production looks a lot like my checklist for people and scripts:

  1. Least privilege. Give an agent the narrowest permissions that let it do its job, and nothing “just in case”.
  2. Policy before action. Check what an agent intends to do against explicit rules before it does it, not afterwards in an audit.
  3. Blast-radius limits. One host, then ten, then a hundred. An agent should be able to change one thing quickly, not everything at once.
  4. A human in the loop for anything irreversible. Deletes, wipes, failovers. Remember Stryker.
  5. A kill switch, and everything logged, so you can explain afterwards what happened and why.

Used well, agents could shorten recovery times dramatically. Used carelessly, they’ll write some very interesting postmortems.

(Usual disclosure: I work for Kyndryl, and these views are my own.)

More Information