Coding agents

A bot can hand a coding job to a coding agent on its Runner: Claude Code, Codex, OpenCode, pi, or Vercel's fx. The bot passes on your request in your words and tells you it started the agent. The agent decides how to do the job, pushes its work, opens a pull request, and reports to you, with the commands it ran and what they printed; the bot then relays that report. You review the work; the bot does not check it for you.

Start one

Ask the bot:

Fix the flaky login test in ~/dev/app and open a pull request.

The bot gives the job a short title, passes on your request with where the repository is, starts the agent, and the chat shows Messaged ⬢ Login: fix the flaky test with the agent's logo. Starting one needs Shell commands in the bot's Access.

The bot adds no plan, checklist, or instructions of its own. The agent works in the repository the bot names, as it would if you started it there in a terminal: it decides how, where, and on which branch, a worktree of its own included, and how to check its work. Unless you say otherwise, it pushes its work and opens a pull request. Anything you want about how, such as keeping your checkout untouched or not opening a pull request, say in your request, and it reaches the agent in your words.

Which agent

Each computer that runs bots keeps its own list of the five, in your order of preference, each with a switch. Find it under Settings → Coding Agents with that computer picked at the top (on the phone, under the computer in Settings). The list shows whether each agent is installed there and its version, and you drag an agent to change the order. A bot uses the first one that is on and installed on its computer, unless you ask for another ("have Codex do it"). Install the agent and sign in to it on that computer, as you would in a terminal.

What it may do

Every command the agent runs goes through the bot's Access and Auto-review, as the bot's own commands do. Commands that only read, and work inside its folder, run at once; a push or a deploy is weighed against what you asked for. When Auto-review asks you, the card under the agent's marker reads "Claude Code wants to run a command" with the command, Allow once, Always allow, and Deny, on any of your devices.

A bot can give an agent secrets it saved for commands, such as NPM_TOKEN to publish a package. The agent has them in its environment; its transcript and messages show their names, never their values.

Follow its work

Click the agent's marker in the chat, or its row under Coding in the chat's details, to open its page: what the bot asked, what it thought and said, the files it explored, the commands it ran, the edits it made with lines added and removed, and its plan, as they happen. The page shows its branch and pull request, and Stop while it works. It is read-only: to change course, tell the bot, which passes your words on to the agent.

The Coding list shows each agent's title and how it stands: Working, Needs attention (it waits for your answer), Ready for review (done, with a pull request), Draft (its pull request is a draft), Merged or Closed (its pull request was merged or closed on GitHub), Finished, Run failed, or Stopped. Agents that worked on the same pull request are one row.

When it is done

When the agent finishes, its report lands in the chat as Message from ⬢ Login: fix the flaky test and notifies you as a reply does: what it did, the commands it ran to check it with what they printed, and its pull request's link. Click it for the whole report and transcript. The bot then relays the report to you in its own reply: the agent's conclusion, the pull request, and the commands it ran with their results, and what the report leaves out. That reply is only a relay: the bot doesn't read the code, build, test, or judge the work, and doesn't send the agent new instructions; if something is missing, tell the bot and it passes that on to the agent. An agent that stops on its own says why the same way, and the bot tells you what happened.

Stop in the chat stops the chat's coding agents along with the bot's work. Quitting Lorca stops the agents at work; one that was done can take another message later and picks up where it left off. The agent's run time counts toward the limits of the turn that started it.