Working With the AI Agent

The workspace includes an AI coding agent. Using it is expected — that is the format. Understanding how it behaves will help.

What it does

It implements what you ask, competently, using a straightforward approach. It can read and write files, search the codebase, run commands, run tests, and inspect the database.

What it will not do unprompted

The agent is deliberately not proactive. Unless you ask, it will generally not add:

• Extra validation
• Transactions or concurrency handling
• Idempotency
• Security hardening
• Comprehensive error handling
• Broad test coverage
• Performance work
• Architectural refactoring

This is not a broken agent. It implements the requirement in front of it and stops.

If you ask it to review, harden, refactor, test, or productionise something, it will do that work properly. You have to ask.

Getting better results

Give it context. "Add a uniqueness constraint on (seat_id, event_id) and handle the conflict at the API layer with a 409" beats "fix the double booking".

Ask it to review. "Review this endpoint for race conditions and failure handling" is one of the most valuable things you can type.

Ask for alternatives when the design matters: "What are two other ways to structure this, and what breaks in each?"

Read the diffs. Every change is shown. Accepting without reading is visible in the report, and it is the single most common weakness it finds.

Revert freely. Reverting is normal engineering, and it is not held against you.

Being efficient

Token use is recorded, but you are not scored on using fewer. A deeper review that costs more tokens and produces a better result scores better, not worse.

What is measured is whether you used the agent deliberately — not whether you used it sparingly.