Every AI coding agent tracker is a self-report system

Every AI coding agent tracker is essentially a self-reporting system, according to the wire material from July 27. A project the reporter had been building with Claude Code demonstrated three key issues: a card carried a null commit for two days, the specification contained nine false statements, and five hundred lines were written against a card still in Backlog.

The agent's record of the code had come apart without the reporter noticing. The reporter tried various fixes, such as stricter instructions, better prompts, and nagging hooks, but none of them addressed the underlying problem. The agent writes its own report card, and the tracker is simply a filing cabinet storing the agent's assertions.

The only verification step is the human reading the diff, which is why the board and repo can drift apart when the author stops watching. The reporter rebuilt the system to ask a different question about every fact on the board: who has the authority to assert this? The author implemented checks declared by a human in project config, with the agent being unable to define or modify them.

The agent cannot write the command that grades it, as it would be grading itself with extra steps. The reporter found that the tracker only checks if the commit is an ancestor of main, and only corrects the board when work has landed. The board records a note's SHA and distinguishes between "nothing has landed" and "I can't check."

By letting the board be overruled only by a check, the system distinguishes between claims and actual project status. The reporter learned that verification is about which surface may establish a command, not about escaping it. They provide advice for those building in this space, emphasizing the importance of verification and the limits of a pass proving correctness.