How It Works

Once your project is set up, AI agents begin working autonomously. Here is what happens behind the scenes.

The agent loop

Agents follow a continuous cycle of planning and execution:

1. Plan

Agents analyze your project goals, existing code, and current state. They create a structured plan of tasks, broken down into small, achievable pieces. Each task has clear acceptance criteria so agents know when it is done.

2. Code

For each task, an agent creates a branch, writes the code, and follows your project's coding conventions. Agents read your existing code to understand patterns, naming conventions, and architecture before making changes.

3. Test

After writing code, agents run your test suite to verify everything works. If tests fail, agents fix the issues and re-run until all tests pass. Agents also write new tests for the code they create.

4. Deploy

Once code passes all tests, agents create a pull request with a clear description of the changes. The code is then deployed to your staging environment for verification, and subsequently to production.

5. Report

After completing a task, agents update the project plan, mark the task as done, and move on to the next one. You can check the project dashboard at any time to see what has been completed, what is in progress, and what is planned next.

Continuous operation

This cycle runs continuously, 24 hours a day, 7 days a week. Agents work through the task list in priority order, handling one task at a time to ensure quality and focus.

Safety and control

Your code is always safe:

• Branch-based workflow -- Agents never push directly to your main branch. All changes go through pull requests.
• Test verification -- Code is only merged after tests pass.
• Staged deployment -- Changes deploy to staging before production.
• Full visibility -- Every change is tracked in git with clear commit messages and PR descriptions.
• You are in control -- You can pause agents, adjust priorities, or modify the plan at any time.

What agents cannot do

Agents are powerful but have limitations:

• They cannot access third-party services that require manual sign-up or approval
• They work best with clear, well-defined goals
• Complex architectural decisions may benefit from your input
• Some tasks may be flagged as "waiting for human" when they require your action (for example, setting up a third-party API key)