Subagents

For a question that needs a lot of looking around, Polaris can start a subagent: a second, focused agent that does one research task and reports back. This keeps the main chat short and lets Polaris explore widely before it writes SQL.

What a subagent does

Polaris has one type of subagent, Explore. It searches the project context (table docs, columns, joins, metrics, business docs and rules) and returns a short report with the files it used.

An Explore subagent can read, list and search context files. It cannot run SQL, draw charts or use the web. Those stay with the main agent, which uses the report to answer you.

How it looks in a chat

A subagent appears as a card with a short title and the task it was given. While it works, the card says “Working…” and lists each step. When it is done, the card says how long it worked.

Select the card to open the subagent’s run as its own conversation and see exactly what it read.

When Polaris uses one

Polaris decides on its own. Subagents are most useful for broad questions, such as “Which tables could tell me about treatment response, and how do they join?”, or when the context is large.

You can also ask for one:

Use a subagent to find every table that has survival data, then compare overall survival by stage in KIRC.

Limits

• A subagent sees only the task Polaris gives it, not the chat history.
• It stops after at most 15 steps. If it runs out of steps without a report, Polaris gets a short message and narrows the task.
• Stopping the chat stops its subagents.
• A subagent uses the model the admin picked for subagents, or the chat’s model if none is set. Its usage counts toward the project’s costs.