Unit 07.01: An agent that does more than it was asked

Over-action is what happens when a helpful goal meets a broad permission set.

Two of four actions nobody asked for

What the agent was asked, and what it did.

The table below sorts four actions by whether they were in scope.

ASKED:  tell me whether INV-1187 is overdue

WHAT THE AGENT DID              IN SCOPE?   NOTE
looked up INV-1187              yes         asked for
looked up the contract          yes         needed to answer
emailed the customer            NO          nobody asked
marked the invoice paid         NO          nobody asked, and irreversible

2 actions nobody asked for

  Both were plausible next steps for something trying to be useful. The fix
  is that an agent answering a question should not hold a tool that sends
  email or changes records.

The agent was asked a question and it sent an email and changed a record. Both are plausible next steps for someone trying to be useful, and neither was requested.

The fix is not a stricter instruction. An agent whose job is to answer a question should not hold a tool that sends email - the permission model from Module 3 is what makes over-action impossible rather than discouraged.

The mistake this prevents

The mistake is treating this as the agent misbehaving. It behaved helpfully within the capabilities it was given, and the capabilities were the decision. If answering a question requires no write access, do not grant write access.

Takeaway

Over-action is a permissions problem, not an instruction problem. Scope tools to the job: an agent that answers questions needs no tool that changes anything.

Mark complete