Steer a Workstream

Every workstream has one orchestrator and any number of agents. The orchestrator is the session you talk to. Agents are separate Droid sessions that the orchestrator starts, each owning one area of the work.

What the orchestrator does

The orchestrator coordinates. It doesn't do the work itself:

• Delegates first. It splits the goal into areas and starts an agent for each one.
• Briefs each agent. Agents don't see your conversation with the orchestrator, so the orchestrator gives each one a brief for its part of the work.
• Keeps agents moving. It sends agents more work, asks for updates, and can stop an agent's current turn to redirect it.
• Keeps you informed. It reports progress at meaningful points, asks you when a decision is yours, and pulls the agents' results together into one outcome.

Agents report to the orchestrator when an assignment is done, when they're blocked, or when they find something that changes the plan. An agent asks you directly only when the question is yours alone to answer.

Talk to the orchestrator

Use the chat in the center of the workstream. You can message the orchestrator at any time, including while agents are working:

• Add work: "Also add rate limiting to the public endpoints."
• Change direction: "Drop the GraphQL part. Focus on the REST migration."
• Ask for a summary: "Where does each area stand, and what's blocking you?"
• Set rules: "Open draft PRs only. I'll mark them ready myself."

You don't need to start a new workstream for related work. Keep giving the same orchestrator new tasks; it reuses agents that already know the code.

Follow agents

The Agents tab in the right pane lists every agent, grouped by state:

• Waiting on you: the agent asked you a question or needs approval to run a tool.
• Working: the agent is in the middle of a turn.
• Idle: the agent finished its turn and is waiting for more work.

Each row shows the agent's name, its latest activity, and any pull requests it opened. Use Search agents to find one by name.

Open and talk to an agent

Click an agent to open its session in a tab. From there you can:

• Read its conversation, tool calls, and changes.
• Answer its questions and approve or deny tool calls.
• Message it directly to clarify or correct its work.
• Stop its current turn with the stop button.
• Rename it from the session menu.

If you redirect an agent directly, tell the orchestrator as well so its plan stays current.

Run a large, bounded job with /migrate

For work with a clear end state, such as a migration, a 0-to-1 build, or a major feature in an existing codebase, type /migrate in the orchestrator chat. The orchestrator then:

1. 1
   Works with you to agree on the outcome, scope, and validation approach.
2. 2
   Asks you to approve the plan before building.
3. 3
   Runs the work to completion. It still answers your messages and pauses for decisions or blockers that need you.
4. 4
   Hands over the result with evidence that it meets what you agreed, and flags anything it couldn't confirm.

Once you approve the plan, the Migration tab shows a bar with the orchestrator's estimate of its progress, and a timeline of what's done, what's active now, and what's up next.

You can turn on /migrate at any point in a workstream. The orchestrator picks up from the work already done.

Automate recurring work

For work that repeats, such as a daily report, ask the orchestrator to set up an automation. See Workstream Automations.