Working with the project agent

Runs and steps

Each requirement you submit becomes a task, and the agent works on it in a run. The steps come from the project’s Harness; with the default Harness they are: understand the requirement, plan the steps, write the PRD, build the HTML prototype, machine check, and collect deliverables for human review.

The Work tab shows the step list and a live event feed for the selected task. The run bar at the top of every page lists projects whose agents are working, and the inbox groups tasks by Running, Waiting for my review, Blocked and Recently completed.

Pause, resume and cancel

You can pause a running task, resume it later, or cancel it from the Work tab. Runs are durable: if the service restarts, an interrupted run is picked up again.

Project knowledge

The Knowledge tab holds facts, decisions, terms and constraints the agent should keep in mind. Add entries yourself, or approve the agent’s learning proposals to add them.

During a run the agent can search project knowledge and read project files when it needs context. It also recalls short summaries of earlier runs in the same project.

Machine checks

Before deliverables reach you they are checked automatically: required sections are present, no template placeholders are left, the prototype is a self-contained HTML page with no external resources, and size limits are respected.

If a check fails, the agent revises the deliverable from the check results. Checks only decide whether work is ready for review; they never mark it complete.

Human review

Every deliverable waits for a person. Approve it, or request changes with a comment. Only an approval completes the task, and the review is recorded in the deliverable’s history.

Language

Write requirements in any language. The agent answers and writes deliverables in the language of the requirement, whatever language the interface is in.