Starting one
Write a prompt in Start, or pick the Agent tile. It starts where you were working, on the runner and in the directory of the column you're on (the main worktree, if that's another), with the coding agent, model and effort set in Settings, Agents, or the ones you picked in Start.
The composer
The chips on the composer change how it works. Model lists every coding agent on the runner and its models, with a search, and the model's effort under them. Access sets what it may do without asking. The rings beside them fill as the context window and your coding agent account's plan usage do; press them for how many tokens are left before the coding agent has to compact, and to pick which of the two the chip's percentage shows. When a turn leaves less than 10k tokens before the coding agent has to compact or stop, it compacts right away instead of when you send the next message; Settings, Agents turns that off.
Access
- Supervised asks before it edits files or runs commands.
- Auto-accept edits edits freely and asks before it runs commands.
- Auto works on its own and asks only when the coding agent flags a risk.
- Full access never asks. Keep it for machines you can rebuild.
Talking to it
Enter sends, Shift+Enter starts a new line. A message sent while it works goes into the turn when it can, or waits for the turn to end; one that couldn't go comes back to the composer, with why under it. / lists commands: /model, /access and /new, then the coding agent's own. @ completes files, folders, other columns and your runners.
Attach files with the paperclip, by pasting or by dropping them: up to 10, 20 MB each. A pasted screenshot comes in as an image.
Select part of a reply and press Comment to quote it: the quote waits in the composer with what you wrote about it, and goes with your next message.
Reading the feed
A file the agent links opens in place, under its paragraph, with the linked lines marked; pick lines in it to add them to the chat. Code blocks, diagrams and pages fold, and any image opens larger.
Points of interest, on the composer, lists the places worth going back to, newest first: your prompts, the reply that ended each turn, what the agent showed, and the code and images in its replies. Pick one and the feed goes there.
Prompts and questions
When it asks, the prompt shows in the feed with the coding agent's own choices, like Allow once or Deny. A question comes with its options, or write your own answer.
Work that goes on
An agent can start work that outlives its turn, like a build or a deploy, and be told when it ends. It can also set itself a wakeup: a message to itself in 20 minutes, at a time, or every so often. Chips on the composer count both; press one to see them, and to stop or cancel one.
Where it works
The bar under its header shows its runner and directory. The directory chip opens folders and branches: switch branch, make one, or make it in a new worktree with Shift+Enter, so two agents can work in one repository without meeting. A new directory starts a new conversation.
Picking another runner on the runner chip moves it there once its turn ends, with its conversation; the agent brings what it needs of its folder. A runner it can't go to says why. The +N -M chip counts its uncommitted changes; press it to open Review beside it.
From an agent
The + on its bar opens columns that start where it works: Review changes, Open terminal here, Open database here, Open canvas here, Debug here and Open editor here. Settle, in its column menu, keeps its conversation for later.