Skip to main content

Multi-Agent Collaboration

In Creating & Configuring Agents you set up the tree. This page is about how that tree actually works when agents start helping each other.

You do not need to master all of this on day one. But it is useful to know what Taurus gives agents, because once you do, you can prompt them to use it properly.

Delegation​

A parent agent sends work to a child with Delegate. It can wait for the result, or fan work out to multiple children and collect results later.

The Supervisor tool​

Supervisor is how a parent manages the team itself: checking who exists, creating or updating children, sending them messages, or stopping work that is no longer needed.

Both Delegate and Supervisor are standard Taurus tools in the agent's tool set.

If a parent wants to peek into a child's work without opening the whole transcript itself, Taurus also gives it Inspect.

Subruns vs. delegation​

A subrun is basically Delegate to self.

The agent spins up a fresh context window for itself, but stays the same agent: same container, same prompt, same tools, same files. If you have used Claude Code or Codex subtasks, it is that kind of move.

A delegate call is different. That sends work to a real child agent with its own context, its own prompt, and its own private workspace.

A good rule of thumb:

  • use subrun when the same agent just needs a fresh head
  • use delegate when the work belongs to a specialist

Example: imagine an agent doing deep research for you. It might use a subrun to analyze a pile of sources without bloating the main conversation. But if you also have a dedicated writing agent or fact-checking agent, then delegation makes more sense.

Mid-run progress​

On long tasks, child agents can send progress updates back to the parent with Message without ending the run.

What's next​

To run agents on a schedule without manually prompting them each time, see Scheduled & Autonomous Agents.