Agents¶

An agent works on one task at a time. It reads the task, claims it so you can see who is on it, and writes its progress back to the task. You can start one from a task, from the composer, or from a terminal.

Ask for work¶

Type a request in the composer at the bottom of any page. Add what it is about:

Tasks: type @ and pick one, or drag a task row onto the composer.
Pictures: paste, drop, or press Picture.

Then send it.

Key	What happens
`↵`	Queue it. An agent picks it up in the background.
`⌘↵`	Send it now and open the conversation.
`⇧↵`	New line.

With no task attached, SBT files one from your first line, named as a short action ("the export button does nothing" becomes Fix export button). With several tasks attached, each gets its own agent.

Start from a task¶

Select a task and press a. Its agent opens in a pane beside the list. If none is running, one starts.

Talk to an agent¶

The conversation shows replies as they stream in, with each tool the agent used and its result.

Type while it works to steer it. ↵ sends at its next step.
⌥↵ queues a message for when it is done.
Stop ends the current run. The conversation stays.
Esc puts the conversation away. The agent keeps working.
Resolve ends the agent and closes the conversation for good. Its history stays with the task.

A message the agent has not read yet shows above the composer with Edit, Cancel and Send now.

See who is working¶

When an agent is working, it shows at the top of the window. Click it to open its conversation.

Agents & jobs lists the agent sessions in your workspaces, including Claude Code and Codex sessions you started yourself in a terminal. Each says where it stands:

State	Meaning
Waiting on you	It asked a question and is paused for the answer
Trouble	Its last reply failed, it stopped moving, or its context is nearly full
Working	In the middle of a turn
Handed back	It finished and said what it did
Landed	It finished and its branch was merged
Quiet	Nothing new for a while

The same page has Scheduled, for jobs that run on a timer, and History, for past runs.

Let an agent drive the window¶

An agent you start from the window can see the page you are on, change the list you are looking at, and open pages for you. Ask it "show me what is in review" and it filters Tasks. Turn this off with the flag desktop_agent_view (see Settings).

Where agents keep files¶

Each conversation gets a dated folder in ~/.switchbard/sessions/, such as 2026-11-09-fix-safari-export. The agent works in its scratch folder and moves a file to deliverables only after you approve it. New versions sit beside old ones (plan.v2.md) instead of replacing them.

Choose the AI¶

SBT's agents run on Pi, which works with the provider you already use: Anthropic, OpenAI, Google and others. You connect one during the first run. An idle conversation stops its agent after 30 minutes to free memory; your next message picks it up where it left off.

You can use Claude Code or Codex instead. Set agent_harness to claude, codex or pi in ~/.switchbard/settings.toml. Claude Code and Codex run in a terminal pane rather than a chat.

Teach your own agent about SBT¶

If you run Claude Code, Codex or Pi yourself, install SBT's skills so they know how to read and update tasks:

sbt skill install --agent all

This writes skill files to each agent's personal skills folder. It adds no hooks and stores no credentials.