Work with an Agent on implementation

Once your intent is clear, set a specific task in chat or use Realize in the sidebar to have an Agent work from intent that has not yet been investigated.

Before starting, make sure the project has an accessible implementation folder, the selected Agent is Ready, and any direct edits to intent have been saved.

Set a task in chat

Start a new chat, choose the Agent, model, and permissions, and use @ to reference a document or paragraph. Describe what you want done:

Update the check-in feature to follow the referenced reward rule. Keep the existing storage approach, check repeated clicks and a 7-day streak, and report the verification results.

A clear goal, scope, and verification request make the result easier to judge. If the task involves a new product decision, you can ask for an intent proposal first and review it before implementation.

Use the sidebar's Realize action

When the project has content awaiting investigation, the sidebar offers Realize. Clicking it opens a new chat and asks the Agent to implement that intent.

Content awaiting investigation includes new text and edited text that has not been investigated again. The status means the current wording lacks its own explanation; it does not prove the feature is missing.

If you only want to find out whether existing code meets a requirement, use Explain. If Realize is unavailable but you have a specific code change in mind, you can still request it in a regular chat.

Continue after accepting a proposal

When an intent proposal still needs implementation work, Accept first applies the document changes, then asks the Agent in the current chat to continue implementing them.

If the proposal states that the implementation already meets it, acceptance records the requirement without starting another implementation run. For example, you may want to formally keep a rule the code already follows.

Implement Directly asks the Agent to implement the proposal without writing it into the intent document. Accept a reviewed proposal when the decision should become a lasting requirement. See Discuss and review intent changes for the full distinction.

Follow progress and handle permission requests

The chat shows the Agent's replies, execution trace, and tool calls. Expand a tool call for details; file edits display their changes. If the Agent provides a plan, its tasks show pending, in-progress, or completed states.

The input's permissions control lets you choose from the selected Agent's permission settings. Options differ between Agents. When asked to approve an operation, inspect its target and scope before deciding.

Stop or continue work

Use the stop button in the input to interrupt the current implementation. It does not undo changes already written to files.

If a proposal's implementation does not finish normally, its card shows Implementation not finished and offers Continue Implementing. In a regular chat, send a new message asking the Agent to inspect what has already been done and continue.

Read errors before retrying. Distinguish a message that was never sent from work the Agent already started, so you do not repeatedly request actions that have partly run.

Check the result

Read the Agent's verification results, then return to the intent document and inspect the explanation and evidence. Request extra checks explicitly when needed, such as duplicate claims, time-zone changes, or failed saves.

If explanations have not updated automatically, run Explain. See Keep explanations up to date for automatic triggers and settings.

Next, read Read explanations and evidence.