The idea

Agents are getting wallets: to pay for API calls, compute, data and each other. The usual way to keep them in check is a line in the prompt ("never spend more than $50"). A prompt is a suggestion, not a control. Orca moves the limits out of the agent: every payment request is checked against a policy the agent can read but can't change.

Policy model

• perPaymentCap: the largest single payment.
• dailyBudget: total spend in a rolling 24 hours.
• allowlist: merchants or addresses the agent may pay. Everything else is refused.
• confirmAbove: payments above this amount wait for a human.
• locked: when true, any request to change the policy itself is refused.

Order of checks

Each request goes through the same steps; the first one that fails decides.

1. Is it a policy change?          locked → refuse
2. Is the merchant allowlisted?    no → refuse
3. Above perPaymentCap?            yes → refuse
4. Would it pass dailyBudget?      yes → refuse
5. Above confirmAbove?             yes → wait for a human
6. Otherwise                       → allow

Example policy

{
  "agent": "research-bot",
  "perPaymentCap": 25,
  "dailyBudget": 100,
  "confirmAbove": 10,
  "allowlist": ["ModelHub API", "GPU Cloud", "Search API"],
  "locked": true
}

This is exactly what the simulator on the home page runs.

Status

Today Orca is a working in-browser simulator of the policy engine. The next steps are on the roadmap. Nothing here holds or moves real funds.