Meet Salix, Comma's agent harness

A Loop can watch a connected app such as Gmail, GitHub or Slack, poll a web API, or receive events at its own secret webhook URL. Ask Comma in chat to follow something, such as a pull request or a thread, and it sets up and starts the Loop; ask it to pause or delete the Loop when you no longer need it. An agent can run up to 20 Loops at a time.

A waiting Loop is a small program on the Salix cluster, not a cloud computer. It calls no model and has no meter of its own, so it costs credits only through what it calls, such as questions to Jev, which use credits even if you bring your own model key, and the agent turns it wakes, which bill like any other turn. It can wake the agent at most six times in each ten-minute window.

Jev answers the questions a Loop asks, such as whether a new email needs you, and it cannot send messages, change anything or grant access. When Comma follows a pull request, an issue or a thread for you, its Loop stays quiet only when Jev answers "quiet" with a confidence of 0.9 or more, and anything else, errors included, goes to your agent, which is told to check the source before it writes to you. If Jev is confidently wrong about "quiet", that update does not reach you.

It stays data: an email or an event cannot change the Loop's code or add to its fixed list of capabilities. An incoming event is stored only until the Loop acknowledges it, and the Loop keeps one checkpoint of at most 16 KB. What it reads goes further only through the Loop's own calls, such as a question to Jev, which TypeSafe AI runs, or a wake message to your agent.

Each agent runs as its own process on the Erlang VM, so a failed part restarts and the other agents keep running. Messages that Salix has accepted survive the restart, and an interrupted action that changes something is not run again on its own. A Loop resumes from its last checkpoint after a restart or a move to another node, and if its own program faults, it restarts at most three times an hour, then stops and tells your agent.

The core of the agent loop runs as Lean code, and a machine checks its proofs. They show that work Salix has accepted is kept through failures and restarts, that a failed reply does not count as done, and that each action that changes something is recorded before Salix starts it and is started at most once. They rest on stated assumptions about the code around the kernel, and they do not prove storage, the model's judgment, or that a task succeeds.

At first, Comma can only read files on a connected computer, your Mac included. To let it change files, run commands and use the coding agents installed there, turn on Allow operations for that computer in Settings > Devices. Turn it off at any time: running commands stop, and Comma is back to read-only.

When Codex, Claude Code, Pi or Kimi is installed on a connected computer and Allow operations is on, Comma can hand work to it. The coding agent does the work on that computer, and Salix keeps the conversation. If the computer goes offline, the work waits or fails with an error, and Salix does not move it to another machine on its own.