Put a seatbelt on your AI agent.

Coding agents are great until the one time they run rm -rf in the wrong directory or drop the production database. This generates the config to wire in infraveil-guard — so destructive commands stop and wait for your approval. The agent can't approve itself.

Cooperative, but un-self-approvable

The viral incidents — an agent that deleted a production database in seconds, backups and all — happened because nothing sat between the model deciding to run a command and the command running. The fix isn't to stop using agents; it's to make the destructive, irreversible actions wait for a human. That's all this does, and it does it by construction: the agent calls the gate, the gate blocks the dangerous ones, and only a human, in their own terminal, can mint the one-time code that lets one through. Every decision lands in a tamper-evident local log.

It's a cooperative guardrail — it works because the agent is told to route through it, not because it traps the agent. For a gate the agent genuinely cannot skip, because it runs inside a governed runtime with central audit across a whole fleet, that's the full Infraveil control plane. This is the seatbelt; that's the system.