Skip to main content
When a task splits into independent pieces, the agent hands them to subagents and works on something else while they run. Each one is a separate job with its own step budget, so a research pass and a repository read can happen at the same time rather than one after the other.

​
The kinds

Research runs shallow or deep. Deep spends more steps and reads more pages. GitHub reads at most 15 files, chosen for relevance rather than crawling everything.

​
Running in parallel

By default two run at once. Fast mode raises that to five, Max mode to four, and attaching a Computer raises it to five as well. The agent can change it for a chat if a task genuinely fans out, up to ten. Guards exist so a loop cannot run away: at most 48 subagents per turn, at most eight left running in the background at any moment, and a kind that fails three times in a row is not tried again. Each subagent gets 32 minutes.

​
Watching them

A subagent appears as a card in the conversation rather than a bare tool call, showing how many actions it has taken and how many tokens it has used. Open it and you get the task it was given, a live activity timeline, any files it touched, and its result or its error. The details panel on a message counts subagents dispatched, web research done and pages read.

​
Background subagents

The agent can start a subagent, carry on with something else, and collect the result later. That is what lets it fire three research questions at once and then write one answer from all three. You do not manage this. It is worth knowing about because it explains why the agent sometimes says it is waiting on something.