Your standup, run by a local brain
Arbiter's local brain now runs the operations loop: a morning briefing across your projects, a tool-calling chat that knows your repos and recent runs, and dispatch of approved work to your coding agents, with an approval gate and a kill switch in front of anything that matters.
The expensive part of working with AI agents isn’t the coding. It’s the supervision: deciding what to do next, kicking off the right job, checking whether the last one finished, noticing the one that wedged. That overhead doesn’t need a frontier model. It needs a model that’s always on, local, and cheap.
That’s what Arbiter’s local brain does now. It runs the operations loop so your frontier tokens go to shipping code, not to “what should I do next?”
A briefing, not a blank prompt
Ask Arbiter for a brief and the local brain reads across your projects and hands you a status summary of where things stand, what moved, and what’s waiting. You start the day pointed at decisions, not at an empty terminal trying to remember where you left off.
A chat loop that knows your world
The brain runs a tool-calling conversation grounded in real signals: your projects, your git state, your recent runs, what you’ve asked it to remember. You talk to it in plain language and it calls the right tools to answer, so “what’s the state of the payments service and is anything stuck?” is a question it can actually act on, not just autocomplete.
Dispatch, with a gate in front
When there’s work to do, Arbiter dispatches it to your coding agents and tracks it as a real, accounted run. Anything irreversible stops at an approval gate first; you approve or deny, and the policy on what needs a human is yours to set and flip. A kill switch and a budget ceiling sit over the whole thing, so the conductor can run ahead without running away.
Where it stands
This runs today in Arbiter’s working alpha: the briefing, the chat loop, and dispatched work are proven against the live daemon on real hardware. It’s still alpha, and we’re hardening the real-world loop before any version promises.
Arbiter is the supervision layer over the rest of the stack: TheAuditor for ground truth, Warden as the agent it dispatches, Curator for memory, BenchProctor for proof. See it move in Arbiter is alive or follow the RSS feed.