Subagents

When you ask for a large app or change, the agent doesn't try to do it all in one leap. It breaks the work into smaller pieces, works through them in order, and shows its progress as it goes. You don't manage any of it directly, but understanding how it works helps you read what's happening during a long build and know when to step in.

Breaking Up the Work

Given a substantial prompt, the agent first decides on the pieces of work it needs to do, then tackles them one at a time. A "build me a habit tracker" prompt might become: set up the screens, add the data store, wire up the add-and-edit flow, then style it. Each piece builds on the last.

This is why Plan mode is useful for big asks. Planning first lets you see and agree on the pieces before the agent starts building them, so the work heads the right way from the start.

Following Progress

You follow a build in the feed in Agent Chat, where the agent shows each step it takes, the files it touches, and the commands it runs. A task list at the top tracks how far along it is (Task 2 of 5 while it works, then 5 of 5 Tasks Completed when it's done), and you can expand it for the full checklist of what's done, working, and still to come.

You don't have to watch every line; the feed is there for detail when you want it. If something looks off partway through, steer the agent or stop it without waiting for the whole build to finish.

Delegating to Specialists

For a large, self-contained piece, the agent can hand the work to a focused specialist, such as a frontend or backend architect, that does that one job and reports back before the main agent carries on. It's automatic: you don't set it up or direct it, and it keeps a big build organized.

In the feed, a specialist shows up as a collapsible block labeled with the piece it's handling, its own progress and steps folded inside. Expand it for the detail, or leave it collapsed and follow the overall progress.

For Longer Builds

A few things follow from how the agent works:

• Big builds take longer and cost more. More pieces means more work, which means more credits. Planning first keeps this efficient.
• Progress is visible, not hidden. You can always see what the agent is up to in the feed, so a long build never feels like a black box.
• You can intervene any time. You don't have to wait for the agent to finish all its pieces. Steer it mid-build or stop it if it's heading the wrong way. See Iterating.
• Smaller asks stay simple. A quick tweak is just done; the agent only breaks work into pieces when the size calls for it.