Work: tasks, routines, missions and arena

Work gathers everything your agents do in the background: one-off tasks, schedules, loops that repeat until a check passes, missions with a clear definition of done, and the Arena, where several coding tools try the same prompt side by side. It also holds Summarize a video, which uses the AI provider.

Where it is

LLM › Work. The second row has All work, Tasks, Routines, Missions, Arena and Summarize a video.

All work

A read-only board of everything in progress: tasks, loops, missions, threads, chat runs, questions and budget stops. It is grouped into Needs you, Running, Waiting to start, Ready for review and Done recently. Each item has a tag for where it came from (Task, Loop, Mission, Thread, Chat, Question or Budget). You cannot create anything here; the empty page links to Create a task.

Tasks

A task runs one prompt in the background with one of your agents. Open its details to see the result, approve a tool or stop it.

1. Open LLM › Work › Tasks and expand Create a task.
2. Pick the Agent.
3. Optionally pick a Workspace with Choose folder…. Without a folder, the agent runs without a workspace.
4. With a folder, choose Where it writes and the Run mode (see the table).
5. Write the Prompt and click Run.

Option	What it does	Default
Where it writes	Automatic: its own worktree only if another job is writing there, Always in a git worktree of its own, or In the folder itself (waits its turn).	Automatic
Run mode	Ask before anything, Auto edits, Auto (CLI reviewer) or Full access.	Auto edits
Jobs at the same time	How many tasks run at once.	4
Jobs per agent	A limit per agent; 0 means no limit.	0

Each task has a time limit of 60 minutes by default; change it in Models and limits. Filter the list with All tasks, In progress, Needs attention and Finished.

Schedules and webhooks

Under Schedules and webhooks you can start a task on a timer or from a webhook.

Option	What it does	Default
Kind	Schedule or Webhook.	Schedule
Cron line	Five fields (minute hour day month weekday), or @hourly, @daily, @weekly, @monthly. Times are local.	*/15 * * * *
Missed while closed	Run once, Run each one (up to 25) or Skip them.	Run once
If the last run is still going	Skip this time or Start another anyway.	Skip this time

Click Create schedule. Schedules only fire while OmniGet is running.

Routines

A routine (the page calls it a loop) runs the same prompt round after round until the work checks out.

1. Open LLM › Work › Routines and expand Create a loop.
2. Pick the Agent, give it a Name, write the Prompt and Choose folder….
3. Set the stop conditions: Max rounds, Max minutes, Budget (US$), Token budget and a Check command (for example npm test). The loop is done when the check passes.
4. Click Review loop, then Start loop (or Edit configuration).

Option	Default
Max rounds	10
Run mode	Auto edits
Resume by itself after a restart	Off

A loop ends as Done, Failed, Cancelled, Blocked or Budget reached, and can be Paused.

Missions

A mission is an objective with a definition of done. It only counts as finished when every required criterion has evidence; an agent saying “done” is not enough.

1. Open LLM › Work › Missions and click New mission.
2. Write the Objective.
3. In Who works on it, choose A bot or A group. For a group you can Plan tasks with the coordinator.
4. Under What counts as done, add criteria: Command, File, Tool result, Rubric or Your check, each Required or Advisory. Start from a preset… fills in common ones.
5. Optionally pick a Folder and a Budget (optional) in Dollars, Tokens, Turns or Minutes.
6. Click Create and start, or turn off Start right away and click Save as draft.

Missions share the same slots as tasks. A mission that is Checking is verifying its criteria; Partly done means some required criteria still lack evidence.

Arena

The Arena sends the same prompt to several coding tools at once, each in its own thread and worktree from the same base, so you can compare the results, vote and apply the winner.

1. Create a project in Threads first. It must be a git repository.
2. Open LLM › Work › Arena.
3. Choose the Project, the Base branch, a Name and the Prompt (or From a thread’s plan).
4. Under Tools, add between 2 and 5 entries, each with an optional model. You can add the same tool twice with different models.
5. Optionally set a Check command with its timeout (default 10 minutes).
6. Click Start in … tools.

When they finish, use Run checks, compare, then Apply winner and Clean up. Stop all stops every entry; Scoreboard shows past results. The Access for the entries defaults to Accept edits.

Summarize a video

1. Set up the AI provider.
2. Open LLM › Work › Summarize a video.
3. Paste the Video URL, pick the Summary style (Short, Balanced or Detailed; default Balanced) and, optionally, the Output language (optional).
4. Click Summarize. The result is saved in AI history.

OmniGet first uses the video’s subtitles, including automatic ones. If there are none, it downloads the audio and sends it for transcription, which only works with OpenAI or an OpenAI-compatible endpoint that offers transcription. Provider charges may apply.

Common problems

“Configure an AI provider first”

Summarize a video uses the AI provider, not your agents’ connections. Set one up and save.

“No transcript available for this video”

The video has no subtitles and the audio could not be transcribed. With Anthropic as the provider, a video without subtitles always ends here.

“This folder is not a git repository: the arena needs one.”

Run git init in the project, commit once, and try again.