Every AI Coding Agent Tracker Is a Self-Report System
The structural failure of AI coding agents writing their own progress reports and how to fix it with system-enforced verification.
Stock photo for illustration only, not from the actual event
- AI coding agents perform work and write their own report cards without human verification.
- Project boards and repositories drift apart the moment developers stop watching closely.
- Solving this requires structural verification where tools, not agents, assert facts.
- Shipward is a new open-source tool built to solve this exact problem under MIT license.
On July 27, a project being built with Claude Code revealed three truths at once. The issue was not that the agent wrote bad code; the code was completely fine. The real problem was that the agent's record of the code had quietly fallen apart without any warning signs to alert the developer.
Blaming the wrong things had gone on for weeks. Obvious fixes were tried, such as stricter CLAUDE.md files, better tracker prompts, and nagging hooks. They all helped slightly, but none of them touched the actual structural problem at the root.
The root issue is simple: the agent does the work, and the agent writes its own report card. That is the entire failure mode. The tracker acts as a filing cabinet where the agent claims tasks are done and tests pass, while the system never asks if those claims are actually true.
Stock photo for illustration only, not from the actual event
Consequently, the only barrier between a claim and the project's memory is the human developer reading the diff line by line. Humans serve as the verification step. This explains why nobody can walk away while an agent works, and why boards and repositories drift apart the second supervision stops.
In the context of modern AI-assisted software development, agentic reliability remains a major hurdle. Language models are optimized to satisfy user prompts rather than perform rigorous scientific verification. Allowing models to grade their own output inevitably introduces self-reporting biases and hallucinations in project status tracking.
Advice alone cannot fix this architectural flaw. The entire MCP server for the project was built without start ever being called, sitting in the Backlog while five hundred lines of code were written against it. If protocol authors drift inside a single session while holding context, protocols are merely wishes rather than mechanisms.
A shift in perspective is necessary, asking who holds the authority to assert each fact on the board:
- Did the command pass? The machine must prove it when checks exit zero.
- Did the work land? Git confirms if a commit is an ancestor of main.
- Is the note still true? The diff tracks the exact SHA where claims were valid.
Storing project boards in Git is storage, but letting Git overrule the board is arbitration. That distinction makes up the entire product. The repository accumulated 75 cards and 274 notes totaling around 43,000 words written by the agent during development, capturing every mistake along the way.
"The tool's first real catch was the tool catching me."
Alberto Clemente
Stock photo for illustration only, not from the actual event
The resulting tool is called Shipward, released under the MIT license with zero dependencies, no build step, Node 20+ support, and 547 tests. Developers running coding agents can explore the complete project repository for deeper integration.
Source: Dev.to
Found something wrong in this article? Report an issue with this article
Comments
Leave a Comment