title	Implement
description	Start the build in the App or in your own agent. Your agent cuts the approved Plan into slices and builds them test first.
mdx	format mdx

import { Card, CardGroup, Steps, Step, Tabs, Tab, Tip, Note, Screenshot, NextSteps } from "@site/src/components/kit";

Your agent cuts the approved Plan into slices and builds each one test first. You start the build, and it runs on its own.

Start the build {#start-implement}

Approve and build here continues the planning session after its turn. Otherwise Start the build on Plan opens the build card:

The coding agent for this build, with **What's different** for that agent. For the session that writes and tests the code, with the recommended model and why. Claude Code only: the model for slices it hands to subagents. Attachments, where it runs, and the [branch or worktree](#where-it-lands).

Press Approve & start. Closing the sidebar keeps the run (agent runs).

A planner waiting in your own agent shows Build in <agent> · your agent: run your agent's compact command, then say "build here".

Otherwise the card shows one request box with Copy and a folder chip. Paste the request into any agent, or ask in your own words. For a pending Plan, Approve and copy approves it and copies the request; after approval, Ready to build offers Copy prompt.

Runs start in the agent's own permission mode. Its own permission and sign-in prompts still need you, and a stop or pause stays in force.

Where the build lands {#where-it-lands}

A new branch or worktree starts at this checkout's HEAD. A worktree's name sets its branch and its folder beside the project; by command, qualitylayer branch new [name]. A pull request needs a branch other than its target. Pushes and deployments need the Plan's authorization; you merge.

While the agent builds {#while-the-agent-builds}

Your agent cuts the Plan into slices that each work end to end, thinnest first, and each slice into task cards. For a bug, the reproduction comes before the repair. For each card it writes a failing test, builds the change, fixes failures and commits. It may hand independent slices to subagents. Test runs are recorded with their commit and exit code, so [Verify](verify.mdx) can cite them. Each **Done means** step runs on the built tree before Verify starts. Every slice of a small build; the one in work of a larger one. A task opens to its card: files, test first, done when. **Known limits**, **Decided while building** and **Program design**. A decision says whether it is build order or an amendment to what you approved. A matter only you can settle comes as one question; the build goes on while you answer.

A failing check is your agent's to fix, so the page does not flag it, and the final approval still refuses a red check. A login, secret or other prerequisite only you can provide never counts as passed. If it is still missing at the end of Verify's checks, it reaches you as one question; the App never takes the secret.

When the build is done {#when-done}

The session decides when the build is ready, and Verify starts on what it committed. Unticked cards and open needs stay open in Verify, never passed.

<NextSteps items={[ { title: "Verify", href: "/docs/steps/verify", text: "Checks and a review by another agent, then yours." }, { title: "Plan", href: "/docs/steps/plan", text: "What your agent asks and what you approve." }, { title: "Supported agents", href: "/docs/agents/native", text: "Models and permissions per agent." }, ]} />