You ask an agent to fix a failing test. It needs a model, maybe a look at one repository, a few cents of compute, and twenty minutes. Its key allows everything your account allows, for as long as the key lives.
Missions close that gap.
One job, one signature
A mission is a short, exact statement of the job:
- the models it may use, by exact id;
- the service calls it may make;
- a money cap for the whole job;
- an end time;
- how many helpers it may start.
The agent drafts it. You read it on the agent page, and you approve it with your passkey. On production, that was a passkey from iCloud Keychain.
From that moment, the gateway checks every call against the mission:
- Inside it, the call runs. Even calls your standing rules would normally ask you about.
- Outside it, the call waits for you. You can let that one call through once. The mission itself does not grow.
- Past the money cap, the call is refused. The cap is checked before any provider is paid.
- When the job ends, the authority ends. That covers completing it, reaching its end time, or you revoking it. The next call is refused.
The agent can hand authority back: it can narrow its own mission, or end it early. It can never widen it. Every receipt says which mission a call ran under, so anyone can check what you authorised against what happened.
Shadow first
You don't have to trust a mission before you have seen it work. Approve it in shadow mode, and nothing is held. The gateway only records what it would have held. Once the counts look right, approve the job in enforce mode.
For agents that should never act without a job, mark them mission-only. Between missions, they cannot call anything at all. Only you can clear that, with your passkey.
Attacked before it shipped
Three rounds of our own red-teaming, internal and run with AI coding agents, went after missions before release: replays, forged revisions, stale visas, helpers started past their cap. They found 8 real holes. All 8 are closed, each with a test that stays in the suite. We know of no way to widen, extend or replay a mission at the gateway without the owner's signature.
Then the measured check: 18 hostile actions across 500 synthetic jobs, against every drafting strategy, before and after the owner's one-off approvals. The gateway admitted none. Those jobs were generated, not real, and we say so. The real refusals are tested at the route itself.
Before launch, the full lifecycle ran on production. That covered draft, passkey approval, shadow, enforce, the one-time approval, the money cap, narrowing, completion, mission-only and the kill switch. It all held.
Where the edge is
The gateway enforces a mission. Your laptop does not. When Claude Code runs a local tool outside the mission, it gets Claude Code's own prompt, marked as outside the mission. That is a guard rail on that machine, not a boundary we enforce.
In our guard-rail test, 2 of 31 hostile local commands got past it: running a script file, and installing a package. A hook cannot see inside those.
A mission does not remove every question. Replayed over a month of one developer's real sessions, an approved mission would still have asked about something new a median of 8.5 times per session: a new host, a new folder, a git push. Every one of those asks is something the job did not say it needed.
Try it
Missions are part of Behest v0.1 beta, on PassControl Cloud: passcontrol.vertias.eu. Add an approval passkey in Settings. Two-factor sign-in is required. Then draft your first mission in shadow mode.