Guardrails in practiceยท

Supervision is not a control

A person watching an AI agent doesn't stop it. Limits set in advance do.

The way teams and developers work, their safety net is someone reading the output from an agent and approving the change.

The trouble is an agent can act before anyone approves it. It can delete a record or send an email, and having someone watching it won't stop it.

I have been there. The agent comes back with half a page of text in the terminal that you have to decipher and work out what change it was working on. It is all too easy to say "yes, go ahead and do that" without understanding the overall consequences.

Many people have spoken about breaking agent-driven development into smaller steps to ensure that the understanding can sit within the mind of the engineer in charge.

This is in direct contradiction of what the AI companies are advising engineers about moving away from the prompts and coding loops to allow the agents to jump through all the hoops they need to get a working system.

All this means we need to get serious about guardrails, consent and permissions.

Amongst other things there need to be clear instructions on:

  • What the agent can touch without asking
  • What tools on a machine each agent has access to
  • Where it must ask for confirmation
  • Clear logs of all agent activities
  • Backups that agents cannot touch
More in Guardrails in practiceAll articles