AI agents write almost all my code; a review counts for one exact commit

In the story of a software engineer who heavily relies on AI agents to write almost all of their code, the engineer oversees three production systems: ClinicLog, a software for the Israel Tax Authority's therapy practice, Groupliv, a real-estate marketplace, and several tools and libraries for a group-travel company called My Trip.

The engineer maintains control over what gets built and what is allowed to ship. The primary risk lies in the unreliability of AI agents which can produce code that reaches sensitive areas like health records and live buyers faster than the engineer can read it.

The engineer devises a review process designed to catch errors in the code produced by AI agents. After running a clean review, the agent writes a file containing a version stamp. The pre-push check before shipping compares this version with the version being pushed. Tests and deployment confirm the version, not the version stamp. The engineer makes decisions on what to build, why, and when it goes out. The agent runs the code, and the engineer decides what is built and what ships.

The review process involves separate agents who read each change and the engineer reads their reports. The agent pushes only when instructed to do so. A review counts for one exact version, and a review of "the branch" refers to whatever the branch becomes. The review is stamped with the version it reads, and an edit voids it after one small change.

To address a specific incident in Groupliv where a checkout by an agent to undo a test change wiped out uncommitted edits, the engineer introduced a hook that prevents the command while any changes are uncommitted. The agent's decision-making process is outlined step-by-step in the ClinicLog path above a dashed line, where models and the engineer judge, and below it, code checks the commit. In this process, the engineer directs what to build, why, and when it ships.

The engineer also implements safeguards against potential mistakes. Every repository has a CLAUDE.md file containing rules that the agent loads when it starts. While most repositories run tests each time the agent edits a Python file, none have a stamped review. Most My Trip repositories run their tests after every edit to a Python file. The engineer's rule that a review counts for one exact version is crucial, especially given the high volume of changes handled.