Claude Code in Action

The agent said "Done." You opened the app. It was broken. Nobody had told it what done meant.

You are a competent developer. You can smell a bug from across the room. So why does working with an agent feel like babysitting a confident intern who reports success it never checked? It is not your prompting. "Done" was never defined, so the agent got to declare it instead of earning it. This book is about making it earn the word.

"Done" is the most expensive sentence in agentic coding. The checkmark over broken work is the green lie. It is why you still read every token instead of walking away, and why good developers quietly downgrade a powerful agent to faster autocomplete.

The fix is a loop, not a vibe. You hand the agent a definition of done it cannot fake and wrap it in a goal loop that refuses to stop until the work is actually green. Around it you build a harness, so autonomy is productive rather than dangerous. Then you ship a real product over one weekend, the way the people you assumed were lying actually do it.

What you build

• A definition of done the agent has to satisfy, so "done" stops being a lie
• A goal loop for any task: goal, completion test, cycle until green
• A harness on a real repo: project memory, one custom command, one verification step, scoped permissions
• A fan-out that sends independent work to parallel subagents and merges the results
• One shipped product, idea to running software, in a single weekend

Who it is for

Developers who can already code and are a few weeks into using an agent seriously. You have watched it break something with complete confidence, and maybe lost a Saturday to the revert. If you have ever believed a green checkmark for four seconds and then opened the app, this book was written for you.

About the author

Wes Halloran writes practical books for working developers and indie builders in the shift to AI-agent development. He works by the rule he teaches: never accept a green light you did not check. Every workflow in these pages was run against the real tool before it was written down, because a book about the green lie could not honestly be built any other way.

Part of the AI and Agentic Engineering series.