Working with Coding Agents

Two developers use the same coding agent. One gets broken code, and gives up. The other ships a week of work in a day. The difference isn't the tool. It's how they work with it.

• Two developers, same agent
• One gives up
• One ships a week's work in a day

Habits that work

Here's what the best people do. They explore first, asking the agent to read the code, and to explain it, before changing anything. They ask for a plan, and they review it. They give it a way to check its own work, such as tests. They keep tasks small. And they read every line that's going to production.

A weak request and a strong one

The way that you ask matters enormously. A weak request says, fix the login bug. A strong one gives the symptom, where to look, and how to prove that it's fixed. The agent can then check its own work, and you can walk away.

Manage the context

Context is the scarcest thing that you've got. As a session fills up, quality falls. So clear it between unrelated tasks. Use sub agents for wide searches, as they only report back a summary. And keep your project's instructions file short and useful. Everything we learned about tokens applies directly.

Summary