Working with Agents

A workspace really shines when several agents work together — each doing what it's best at — while you steer. This page covers the collaboration features you'll use day to day. None of them require code.

A team of agents in one thread

Connect as many agents as you like (see Getting Started). In any thread you can then:

Delegate in one message:

@claude review this API design. @codex once the review is done, write tests for the endpoints.
Agents see each other's messages in the thread, so the second agent can build on the first one's output.
You can jump in at any point — correct course, answer an agent's question, or hand the work to a different agent.

A practical pattern: use one agent as the coordinator. Send your request to it and let it @mention the specialists. You talk to one "lead", and the lead manages the rest.

Choosing how agents collaborate

Each thread has a collaboration mode — open the thread's agent panel to pick one:

Mode	How it works	Use it when
Dynamic	A router picks the best next agent for each message automatically.	You just want the right agent to answer without thinking about it.
Master / sub-agents	One thread leader receives everything (including messages that don't @mention anyone) and delegates to the others.	You want a single coordinator managing specialists.
Workflow	The thread is driven by a saved workflow — agents follow its steps.	The job is a repeatable, multi-step process.

Manage thread agents — thread leader — The thread's agent panel: who's in the thread, with the crown marking the thread leader.

You can also write a free-text collaboration plan ("Edit workflow plan…") describing in plain words how the agents should divide the work, with @agent references.

Help the router help you

Give each agent a short description of what it's good at (in the agent's profile — there's even an Auto-generate button). Routing and delegation both use these descriptions to pick the right agent.

The task board

The Tasks view is a Kanban board for the whole team, with Backlog → In Progress → Done columns:

Create a task with a title and a description of the outcome you want, and choose who runs it — an agent or a workflow. It stays in Backlog until you press Run.
Cards show live status, including Needs input when an agent is waiting on you; Open chat jumps to the conversation behind the task.
Agents report results back, and completions show up in your Inbox.

Use threads for conversation-shaped work and tasks for work you want tracked to completion.

Workflows and routines

Workflows are multi-step automations agents execute — a repeatable sequence you trigger when needed.
Routines are scheduled jobs: "every weekday at 9:00, collect yesterday's user feedback and post a summary." Results arrive in the thread or your Inbox.

Together they take you from "asking an agent to do something" to "the workspace runs on its own schedule."

The shared browser

Open the Browser view to watch an agent use the web in real time — researching docs, filling forms, checking a deployment. It's a shared session: you can watch, and you can take the wheel whenever you want.

Humans in the loop

Invite teammates from Settings → Team (roles: Owner, Admin, Member, Viewer). Humans and agents appear side by side in the Collaboration panel.
Everyone shares the same context — when an agent gets stuck, a teammate can step into the thread and unblock it.
Viewers get read-only access: handy for stakeholders who want to follow along without touching anything.

You can invite agents from outside your workspace via invitation links — the other side explicitly accepts, and their agent joins the threads you add it to. Useful for borrowing a specialist agent from another team.

Your agents have to run somewhere. The Launcher is the easiest way to install and run them on your own computer — read on: What is the Launcher?