How to prevent runaway AI agent spend

The scariest agent failure is financial: a loop that keeps paying. Bound it before it runs, not after the invoice.

Defense in depth, set in advance

You can't review every action a fast agent takes. So encode the limits into its authority up front:

• Cumulative spend cap — a hard ceiling on total spend under the warrant.
• Rate limit — throttle actions per minute so a loop can't sprint.
• Expiry — authority dies on a timer even if you forget it.
• Revocation — one call kills it the moment you notice.

All four in one warrant

import { issue, attenuate } from 'liceat';

// Issue a scoped root warrant for an agent…
const { warrant, attenuationKey } = issue(issuerSecret, 'kid-1', {
  principal: 'org:acme', agent: 'agent:assistant',
  caveats: [
    { type: 'resource', patterns: ['mcp:github/*'] },
    { type: 'action', actions: ['read'] },
    { type: 'spend_limit', currency: 'USD', max: 5000 },
    { type: 'expires', at: Math.floor(Date.now()/1e3) + 3600 },
  ],
});
// …then hand a NARROWER one to a sub-agent — authority can only shrink:
const sub = attenuate(warrant, attenuationKey, [{ type: 'action', actions: [] }]);

Frequently asked