Agent Spend Guardrails

Agentic checkout and agent-to-agent payment rails (x402, AP2, and the wave behind them) are about to let AI agents spend real money on your behalf. The moment that happens, you want controls — before the agent acts, not regret after the wrong $800 flight is booked.

Agent Spend Guardrails turns loose intent ("don't overspend") into an explicit, enforceable spend policy: per-transaction caps, rolling daily/weekly/monthly caps, category whitelists, merchant whitelists, and an approval-threshold ladder that decides what the agent may do alone versus what needs a human yes. It then audits what your agent bought — or almost bought — against that policy and returns a clear per-transaction verdict: APPROVE, HOLD, or BLOCK, each with the binding rule and a one-line reason.

It is built to fail safe: anything ambiguous defaults to HOLD or BLOCK, never auto-approve. New merchants, off-whitelist categories, foreign currency, non-refundable charges, and cap proximity all escalate toward human review. The output is a compact policy block, a verdict table, a short remediation list, and a one-line headline you can screenshot.

The package is a lean SKILL.md orchestrator plus a bundled REFERENCE.md containing the full rule schema, three approval-ladder templates (conservative/balanced/permissive), a trigger catalog, worked examples, and a ready-to-fill report template.

This skill defines and audits policy only. It never executes, authorizes, or moves money, holds no payment credentials, and is decision support — not financial advice. A human owns every real-money decision.