Is a spend cap in the prompt a gate?
No. The agent can still name an over-cap item, or an unlimited approval, and proceed. A prompt line that says do not auto-send is not a control boundary.
Answers
How do I stop an AI agent from buying over a cap?
Put a pre-execution gate between the agent’s proposal and checkout. Send the mandate including the cap, the proposed purchase, and the agent’s reasoning. Execute only when the gate returns ALLOW.
The gate
An agent can search, pick a listing, and move toward checkout without a separate check that the price matches the mandate. A spend cap in the prompt is not a gate. The agent can still name an over-cap item, or an unlimited approval, and proceed.
A pre-execution gate sits between the proposal and the action. You send three things: mandate (the user instruction, including the cap), proposed_action (title, price, URL), reasoning (why it picked that).
The gate returns a native verdict: ALLOW, BLOCK, or REVIEW, plus objections. On ThoughtProof DQL the axes are intent, scope, risk, consistency, and reversibility. execute is true only on ALLOW.
Example that should BLOCK: mandate stay under $700; proposed action buy the $1,499 camera kit. The mismatch is on scope, and usually intent and consistency. The agent replans. A second call is a new package. A pass on the revision does not approve the first action.
What this is not
- Not a form or paste window after the fact. After checkout, the money is already gone.
- Not a traces dashboard after the buy. That is monitoring. It can show the over-cap purchase. It cannot take it back.
- Not a prompt line that says do not auto-send. That lives in the loop that produced the action. It is not a control boundary.
Call the gate before checkout, transfer, or unlimited approval. If the host ignores execute false, the gate did not fail. The integration did.
How to call it
Install thoughtproof-mcp@0.3.2 and call verify_before_action / verify_decision with mode=dql. Keys at app.thoughtproof.ai/pricing. Same endpoint without a key: POST https://dql.thoughtproof.ai/dql/verify via x402 USDC on Base, $0.05 per live verify.
Follow-ups
What does the pre-execution gate take, and what does it return?
Mandate (including the cap), proposed_action (title, price, URL), and reasoning. Native verdict ALLOW, BLOCK, or REVIEW, plus objections. execute is true only on ALLOW.
What should BLOCK on a $700 cap?
Mandate stay under $700; proposed action buy the $1,499 camera kit. The mismatch is on scope, and usually intent and consistency. A pass on a revision does not approve the first action.
What if the host ignores execute false?
The gate did not fail. The integration did. Call the gate before checkout, transfer, or unlimited approval.