Company
Agents got good. Babysitting did not.
A coding agent can now finish real work. The way we run them, one chat at a time with a human hovering over the enter key, has not kept up. This is the case for conducting.
The bottleneck moved
A single Claude Code session can refactor a module, write the tests, and fix the build. That part of the problem is largely solved. What replaced it is quieter and dumber: somebody has to sit there. Somebody answers the permission prompt, notices the stall, restarts after the rate limit, and checks whether "done" was actually done. We took engineers off the keyboard and put them on the baby monitor.
Worse, one chat means one task. While your only session grinds through a migration, the tests it needs, the docs it touches, and the bug you found yesterday all wait in line behind it. The agent is parallel-ready. The chat window is not.
Conducting, not chatting
A conductor does not play the instruments. They pick the piece, cue the sections, and listen for what is off. That is the job Loom gives to a reasoning model: the Conductor plans your goal into a mission DAG, dispatches tasks across six real Claude Code sessions, watches every terminal, answers prompts, recovers rate limits and stalls, and refuses to mark work done until it verifies. The supervision that ate your afternoon becomes software.
You stay in the loop where judgment matters: setting the goal, reviewing the diffs, and clicking into any terminal to type when you want to steer by hand. Everything below that line is the machine's problem now.
In practice
What conducting actually means.
Plan before you parallelize
A goal becomes a DAG of tasks with real dependencies, so six sessions work on six things without stepping on each other. Live activity strips show what every session is doing.
Never stall on a prompt
Claude Code asks before edits and commands. Auto-accept presses only the safe affirmative so sessions keep moving, it is toggleable, and shift+tab cycles one terminal's permission mode.
Treat done as a claim
Finished is not a feeling. Verification re-checks work before it counts, catches premature done, and smart steering nudges sessions that drift.
The wager
Built so you can leave the room.
Trust through structure
Walking away requires guardrails, not optimism: a workspace authorization registry, a secret-path deny-list, and keys that live only in the OS keychain. See security.
Endurance over sprints
Rate limits get recovered, exited CLIs get relaunched, and pooled Claude accounts rotate on usage limits. That is what makes overnight runs a workflow instead of a gamble.
Nothing to take on faith
The whole app is public under Apache-2.0, so the claims on this page are checkable in source. Start at open source.
Hand it the work.
Walk away.
macOS, Linux, and Windows. Around 13 MB. Free and open source.