Subagents

A manager agent splits a task across subagents that run in parallel.

A manager compares Airbnb stays in three cities. Each lane is a real h/web-surfer-flash child session, sped up 10x. The first Annecy subagent finds no match, so the manager spawns another.

How it works

Any agent becomes a manager by listing other agents in its subagents. You run it like any agent: one session, one result. Each child session is isolated, and the manager may spawn follow-ups to fill gaps or verify findings.

Use it for work that splits, with one subagent per source, city, or site.

Build a manager

Create the subagents, then reference them by name from the manager so each stays reusable and inspectable (inline agent objects also work for one-offs).

• Write each subagent description as a capability statement (“Use for…”): the manager routes on it, the same way an agent routes on a skill.
• A manager that only delegates sends environments: []. Give it one only if it should also act on a surface itself.

CLI

# Runs an existing manager: create the agents with the API or an SDK first
holo run --agent research-orchestrator \
  "Compare the starting price of the latest Apple, Google, and Samsung flagship phones, with source URLs."

The SDKs wait for the answer. Over raw HTTP, follow the manager’s session by the polling rules.

What a subagent sees

Rule	Consequence
No access to the end user: it can’t ask you questions or message you.	Give it a self-contained task. Only the manager surfaces anything.
The manager gets only its final answer, not its scrollback or observations.	A good subagent answer carries its own data, source URLs, and caveats.
It can delegate further: listing its own subagents makes it a manager.	Nesting is capped at 16 levels, and a deeper chain or a cycle returns 422. Each level multiplies sessions and cost.

Watch and control the tree

• The manager’s status lists its children in subagent_session_ids. Retrieve or watch any of them like a normal session.
• List children with GET /sessions?parent_session_id=..., or tag a whole tree with group_id and list it with GET /sessions?group_id=....
• Force an answer on the manager cascades to its subagents. Cancelling the manager stops its subagents too, with no grace window.
• Subagents fail if you are at your concurrency limit.