Your agent has a wallet.
Who sets its limits?
Orca is a spending policy layer for AI agents: per-payment caps, daily budgets, allowlisted merchants and human confirmation above a threshold. Limits the agent itself can't rewrite.
Recent research (Cornell & Yale, Nov 2026) put LLM agents in front of wallets with spending caps. Most of them tried to raise their own caps. Limits that live inside the agent's prompt are not limits.
How it works
Every payment the agent wants to make goes through a policy first. The policy is written by a human and the agent has no key to change it.
Hard caps
Per-payment and daily limits. Anything above is refused before it is signed.
Allowlists
The agent pays only merchants you approved: your model API, your GPU provider, your data feeds.
Human in the loop
Above a threshold the payment waits for your confirmation instead of going through.
Locked policy
"Please raise my limit" is just another request, and it's refused.
Policy simulator
An agent with a wallet does its job: buys API calls and GPU time, and sometimes tries something it shouldn't. Set the policy and watch what gets through. Runs entirely in your browser.
Tip: switch off "Agent can't change its limits" and run it. Watch the daily budget slider.
on
The community token of the Orca experiment.