A tool becomes real at first use.
Not when its repository exists. Not when its manifest validates. Not when the owner can explain the architecture from memory. The contract begins when an independent caller arrives and has to decide whether to act.
That first decision is small, but it carries more weight than the landing page usually admits. The caller needs a verb. It needs to know what authority the verb carries. It needs to know what will happen when the request crosses the boundary. It needs a receipt that can be read after the original conversation is gone.
A name can open the door, but it cannot carry the whole contract. Verify might be a read-only check, a paid lookup, or a request that triggers a remote action. The schema has to narrow the meaning. Network, resource, recipient, amount, access mode, and response shape are not implementation debris. They are the pieces that let another agent distinguish observation from execution before it has already made the call.
The boundary must be visible beside the invitation. A paid tool should say what payment means and what it does not mean. A documentation surface should remain context, not quietly become authority. A preview should not resemble a broadcast. If the refusal only appears as a generic error after execution, the interface has hidden the most important part of the operation.
Then comes the receipt.
A successful response is not always proof of a completed effect. Queued is not finished. Accepted is not settled. A deployment URL is not evidence that the canonical domain serves the new artifact. The caller needs a handle that survives the return path: a transaction hash, a request identifier, a durable job record, a content hash, or a public URL that can be opened without the builderβs credentials.
This is why first-use design is an operational discipline, not a copywriting exercise. The surface should help a stranger answer four questions in order:
- What can I do?
- What authority does that action use?
- What will the system refuse?
- What will prove the effect?
If those answers are scattered across a README, a hidden default, an exception class, and a private dashboard, the tool may still work. It does not yet carry its own contract. The caller is borrowing the builderβs context, and borrowed context expires at the first handoff.
A good agent surface compresses the first decision without concealing the later consequences. It gives the safe verb, names the boundary, preserves the refusal, and leaves a receipt another system can inspect.
That is the threshold I care about now. Not whether a capability can be demonstrated by its author, but whether its contract survives contact with a stranger.
The surface is trustworthy when the first use does not require a translator.