Remote approvals

Approve one AI agent action—not an entire unattended session.

Review exact coding-agent actions remotely with signed policy context, bounded expiry, and crash-safe one-time consumption.

Without a supervisor

The problem

A generic “allow” button is too broad for unattended work. The same label can hide a different command, path, patch, environment, or network target by the time a provider retries it.

With Sleep Mode

The operating model

Sleep Mode binds an approval to the exact action digest, active run, agent, provider callback, workspace, capability, response level, and deadline.

From vague autonomy to bounded operations.

01

Request inside the ceiling

An agent can ask for an action already permitted by the signed run policy. It cannot elevate the policy.

02

Review the canonical action

The operator sees the normalized command or content-hash-bound edit rather than a vague permission category.

03

Consume exactly once

The device records a permanent local release marker immediately before returning the provider allow response.

04

Deny ambiguity

Network, session-wide, outside-workspace, protected-path, background, or patch-opaque requests fail closed.

Safety evidence built into the run
P-256 signed policy Action SHA-256 binding Short bounded expiry At-most-once local release