How the agent works

How the agent decides whether to answer, plan, or build, and how sprints and effort shape the work.

There is no mode to pick before you send a request. The agent reads each message and decides whether to answer, ask for context, implement directly, or plan first. Two controls shape the work itself: sprints define a unit of work, and effort sets how deeply a sprint approaches it. Model selection chooses the models used for that work.

How the agent decides

Your message	What the agent does
A question	Answers without changing code
A focused change	Implements it directly in the current session
A larger product or feature	Can plan requirements, user flows, roadmap, and architecture before it builds

To plan first, say so in your message, for example “plan the architecture before you build”. Review the plan in Plan, then ask for implementation. Planning never starts a build on its own.

To keep working on the plan, turn on plan mode with /plan, in the CLI or the web workspace. Each message then changes the plan (flows, spec, roadmap, architecture) and no code runs; on the web, a Plan updated card lists what changed, and Build it starts building. Plan mode stays on until /auto, or until you turn off the plan chip in the web composer. It works on every plan, and each re-plan uses credits like any other message.

Choose a unit of work

For a small change, describe it in chat. Use a dedicated sprint for a defined feature:

/sprint add CSV export to the reports page with a progress indicator

A sprint runs in a new session. In projects with a roadmap, the sprint controls can build the next planned task or continue through the roadmap on Autopilot.

Follow the session that owns the run. Parallel work, queuing, and admission limits depend on the client and account; sending another message does not guarantee that it starts immediately.

Sprint effort

Set effort with /effort before launching the sprint.

Setting	Approach
auto	Choose an approach for each sprint (recommended)
low	A fast direct loop for focused changes
medium	A direct loop organized around a task list
high	A full research, coding, and verification pipeline

Higher effort can involve more model work and credits. Effort controls the workflow; it does not promise a particular duration or result.

Review the outcome

Builds that change code end with an independent verification of your acceptance criteria. State the checks you need in the request, for example type checking, a specific test suite, or a browser journey, and read the diff before shipping.