Agent Team
One prompt. Many hands. Your agent becomes the lead, and Seedmux gives it a team.
With Agent Team, the agent you're already talking to can bring in other agents. It opens them as real panes next to yours, gives each a written task, waits for each to hand in a receipt, checks the work, and comes back to you with one answer.
You don't learn any commands. You just ask.
Split this migration across the team: codex does the schema diff, grok rewrites the docs, and you fix the callers.
What happens
- Your agent plans the work. It writes each task down — the goal, the context the worker needs, what to deliver and how it will be checked.
- Workers appear beside you. Seedmux opens a pane for each worker in the same tab. Your agent keeps half the screen; workers stack in a column on the right. Each pane's title is the task in plain words, like Rename hello to hi in hello.txt.
- Workers do the job and report back. Every worker must hand in a receipt — done, failed or blocked — with the evidence behind it. Failure is reported too; nothing is left hanging.
- Your agent reviews. It reads each receipt, checks the evidence against the task, and sends work back if needed. When several tasks belong to one goal, it waits for all of them and gives you a single, merged answer.
- The workspace tidies itself. Once the batch is delivered, the worker panes it opened are closed. Panes you opened yourself are never touched.
You can watch every worker in its own pane, click in and type, or answer its permission prompts — it's a normal agent session. On your iPhone, workers show up like any other session.
Two ways to work
Dispatch — split the work. Independent pieces go to different agents in parallel: code changes, reviews, research. Use it when the work is big enough that more hands finish it sooner, or when a task suits a particular agent.
Roundtable — get independent opinions. For decisions with real uncertainty, your agent writes a brief and asks two or more agents to answer it independently, without seeing each other's answers. It then summarizes where they agree, where they don't and what evidence would settle it — and follows up only where a disagreement still matters.
Who can join
Workers can be Claude Code, Codex, Grok, OpenCode, Kimi, Antigravity or Devin — whatever you have installed. Name the agent you want, or let your agent choose.
You can also ask for a particular model, or for Claude and Codex, a reasoning effort:
Open a codex with xhigh effort to track down the flaky test.
Smart reuse
Your agent reuses an idle worker when it already has the right context, and starts a fresh one when the old one's context is too full or full of unrelated history. Every task and result is written down, so a new worker loses nothing.
Workers that have sat idle for half an hour close on their own.
Settings
In Settings › Agent Team:
- Enable Agent Team — turn the whole feature on or off.
- Keep worker panes after the batch is delivered — if you'd rather close them yourself.
- Instructions — the guidance your lead agent follows is plain text you can read and edit to taste.
Next: Things to ask for →