Sources checked
Make each action and its inputs explicit
An original calendar tool should make time zone, attendee list, and draft-versus-send behavior explicit. Prefer a focused action such as creating a draft event over a broad command that can silently update unrelated records.
Return enough information to verify completion
Readable output is useful only when it preserves the facts needed for the next step. Include stable record identifiers, useful errors, and whether a requested action actually completed.
Keep the blast radius in the tool, not in the prompt
Assume the model will call the tool with the wrong id once.
The tool should refuse to act on a record the current user cannot see, and it should refuse bulk actions unless the schema has an explicit list with a maximum length.
A description that says “do not delete” is weaker than an API that has no delete method.
Version the schema. When you add a required argument, old prompts fail in a visible way.
When you silently change the meaning of an argument, old prompts succeed and do the wrong thing. Log the arguments and the result id.
You will need them the first time a user says the agent did something they did not ask for.
What to record before you rely on it
For every tool, write one sentence a stranger could use to predict the side effect. Add a negative test: a call that must be rejected.
Store the schema version with the transcript. If you cannot explain the tool without saying “it depends,” split it.
Keep the model id, the date, and the inputs you used beside that note, so a later change can be compared with the same test. When the result is a failure, store the artifact and change the design.
Do not relax the check until the model happens to pass. Name a person who owns that failure.
A note with no owner and no inputs cannot be repeated, and a pass with no date cannot be compared with the next model id.
Common questions
Should one tool wrap our whole API?
No. A wrapper with dozens of optional fields is a place for the model to guess. Expose the few operations the task needs, with narrow arguments.
Do examples in the tool description help?
One accurate example helps. A pile of examples becomes a second prompt and can contradict the schema.
Prefer the schema, and test calls from a real model rather than from the example you wrote.
Sources & further reading
How this story was made
Written by Kristian Kostov with AI assistance and checked against the linked sources. Company performance claims are attributed to the company. Analysis reflects AiLookout’s interpretation; we have not independently tested the products discussed. Cover photography is illustrative and does not depict the specific announcement or product.
Our editorial standards





