Approve Your Agent's Plan From Your Phone, Then Let It Run
With a supported remote channel and approval policy configured, review an agent plan from your phone, approve sensitive execution with a single-use code, and follow the session.
There is a gap between reading what an agent intends to do and letting it do it. Reading is safe. Doing is the part that touches your code. Until now, closing that gap meant sitting at the machine. The working alpha can close it from your phone when the relevant channel and approval policy are configured. You can list your live sessions, have the brain produce a plan, review that plan on the screen in your hand, approve execution with a single-use code, and watch the output land. The whole loop runs from a chat message, and nothing touches your code until you say so.
Look before you approve
Start with /agents to see what is actually running. You get your live sessions listed, so you are
steering real work, not guessing at it. Point /plan at the session you care about by id and the
brain produces a plan for it. That step is read-only and free. It reads and reasons; it does not run
anything. You get to see the shape of the work before a single command executes, which is exactly the
information you want before you commit to it.
The gate sits between the plan and the doing
A plan you like still does not run on its own when the approval policy requires a gate. To let it proceed you send a single-use code from your authenticator app, the kind any standard authenticator produces, and that approval clears the way for the work you just reviewed. Read the plan for nothing, pay a second factor only when you are ready to let it act. The split is deliberate: the cheap, safe part stays open, and the gate lands precisely on the part that changes your machine.
Steer one session, then watch it
Everything targets a specific session by id, so you are never spraying commands at the wrong worker.
Send it more direction with /send, approve and kick off execution with /exec, and follow the
output live with /tail. You approve the plan, it runs, and the results stream back to you where you
are standing. You are not waiting to get home to find out whether it worked. The session reports as it
goes.
The whole loop, off the machine
The pieces exist as a working alpha path. With Telegram and the relevant approval policy configured, plan, approve, execute, and watch can work as one continuous loop with no window to open. You can be away from the machine for the entire thing: see the plan, decide, gate it with your authenticator, and stay with the run until it finishes. The operation follows you instead of pinning you to a desk.
Why the control loop matters
Remote control is only useful if it covers the moment that counts, and the moment that counts is when work starts touching your code. Reading a plan from anywhere was always going to be easy. Approving execution from anywhere, with a real second factor standing between the plan and the doing, is the hard and honest version. That is the behavior the working alpha is preparing for public binary release.
Arbiter is a proprietary working alpha in public binary release preparation. Follow the public release channel or RSS feed for updates.