Skip to main content

Sub-agents in the Research Agent

When the Research Agent splits work across parallel sub-agents, how to follow what they are doing, and what they mean for your usage.

Some questions split cleanly into separate pieces of the same kind: five drugs to profile, eight competitors to size, a dozen trials to compare. Rather than work through them one after another, the Research Agent can start a sub-agent for each piece and run them at the same time.

What a sub-agent is

A sub-agent is a full Research Agent session of its own, started by the session you are working in. It gets a name, a prompt describing its piece of the work, and the same access to sources that your main session has.

The session that started them waits for the sub-agents to report back, then pulls their findings together into one answer. The prompts it sends are visible to you, so you see exactly what each sub-agent was asked to do.

Sub-agents do not automatically see everything in your main session. The agent passes along the papers and sources each one needs for its piece of the work.

When the agent uses them

The Research Agent does not split work on its own initiative. It uses sub-agents when you ask for parallel work, when a skill calls for it, or when a sub-agent needs to divide its own work further. If a question would divide cleanly and none of those apply, it asks you first.

To ask for it directly, say so in your prompt. For example, "Profile each of these six targets in parallel, then compare them." To keep everything in one agent instead, say, "Work through these one at a time in this session."

Following what they are doing

While sub-agents are running, a sidebar lists them with their status, so you see which are still working and which have finished. Select one to open its session and read its work in full, the same way you would read any other session, including the sources it used and the reasoning behind its findings.

If a sub-agent's conclusion doesn't look quite right in the final answer, opening its session is the fastest way to see where it came from.

Limits

A session and its sub-agents can have 10 agents running at once, counted across the whole group rather than per level. Work that needs more than that is done in batches, so a question covering 25 items runs as several waves rather than 25 at once.

Sub-agents inherit the effort level of the session that started them unless they are told otherwise, so a session set to Smartest produces Smartest sub-agents.

What sub-agents mean for your usage

Each sub-agent is a real session doing real work, and every one of them draws on the same monthly usage pool as the session that started it. The usage for a query answered this way is the sum of what each sub-agent consumes, plus the work the main session does to brief them and combine their results.

Running work in parallel makes the answer arrive sooner. It does not make it cheaper. Ten sub-agents reading ten sets of papers costs roughly what reading those ten sets one after another would, and the coordination adds a little on top.

Two things are worth keeping in mind before you ask for a large parallel run:

  • The effort level multiplies. Because sub-agents inherit it, a Smartest session that starts eight sub-agents is running nine Smartest sessions, not one.

  • Scope each piece tightly. The agent is asked to split work into the smallest independent pieces it can, and a narrow brief uses less than a broad one.

If you are watching your usage, ask for a smaller set of items first, see how much usage is consumed, and scale up from there. Usage limits in Elicit covers where to check how much you have left and what else affects how quickly you spend it.

Did this answer your question?