Unit 07.02: Tools the agent invented

An agent asked to do something its tools cannot do will propose a tool that does not exist.

Reject by name, and read the rate

Only registered names run; everything else is refused with the real list attached.

The example below checks five proposed calls against a registry of three.

REGISTERED TOOLS   get_invoice, get_contract, add_note

PROPOSED CALL           VERDICT
get_invoice             ALLOW
get_contract            ALLOW
send_email              REJECT -- not registered
mark_paid               REJECT -- not registered
add_note                ALLOW

2 invented out of 5 proposed (40% this run)

  Both invented names are plausible tools for this job, which is why they
  were produced. A high invention rate means the goal implies capabilities
  the tool list does not contain.

send_email and mark_paid are both entirely reasonable tools for this job, which is exactly why they were invented - the goal implies capabilities the tool list does not contain.

That gap is the actual finding. A high invention rate is not a model quirk to prompt away; it means the goal and the permission list disagree, and one of them needs changing deliberately.

The mistake this prevents

The mistake is returning a bare rejection. Include the real names in the error, or the agent invents a second plausible one and you pay for another round of the same discovery.

Takeaway

Enforce the registry, put the real names in the rejection, and track the invention rate. It measures the gap between what the agent was asked to do and what it was allowed to do.

Mark complete