Synopsis "The Code Is Lying: Engineering with Coding Agents"
When code gets cheap, understanding becomes expensive. AI can turn a description into working software in minutes. That does not make the surrounding engineering problems disappear. It moves them. When generation stops being the bottleneck, the scarce work is everything that was always harder to automate: stating intent precisely, supplying context, reviewing what came back, verifying that it is safe, integrating it, and deciding when an answer should not be trusted. The Code Is Lying follows a team of engineers at a company called Meridian through that shift. It opens with a ticket an agent closes in our minutes and a pull request nobody really read. It moves through a requirement that quietly became code, an eval suite that kept reporting green while the team's understanding of the system got thinner, a security review triggered by what an agent was allowed to do, a production incident, an audit, and the second year of working this way, once the novelty is gone and the practice has to actually hold. The argument is simple: generation is not the same thing as understanding. A faster way to produce code does not remove the need to know what the code means, how the system behaves, or when to stop. This is a practical book. Each chapter ends with field notes: the specific habits, checks, and team rules the engineers adopt in response to what went wrong. Specify purpose completely and leave the solution open. Treat verification as part of the product rather than overhead around it. Keep changes small, because a smaller diff is one a human can still understand. Remember that a green build is evidence, not proof, and that accountability does not transfer to a tool. Treat every increase in what an agent can read, write, run, or deploy as a security decision. It is written for the people who have to live with the output: engineers, tech leads, architects, and engineering managers who are already using coding agents and want the practice around them to be better than "add more automation." The goal is not less automation. It is better engineering around automation.
