AI Pulse · Under the Radar

A small open-source tool called AgentTrail is taking a practical approach to a growing problem with AI coding agents. Agents are good at narrating progress. They can say a task is complete, list files they supposedly changed, and describe a fix with confidence. But that report isn’t always the same as what actually happened in the codebase.

AgentTrail watches the local filesystem and compares real changes against an agent’s declared plan.md. If the agent says it finished a job but the file diff tells a different story, the tool surfaces the gap. It runs locally, uses no cloud connection, and is designed as a passive observer rather than a hard approval gate. The project is only four days old, is attracting about 55 GitHub stars a day, and its creator says it has been tested on repositories with up to 78,000 files.

For a business owner, the point isn’t this specific 470-line tool. It’s the operating model behind it. As more teams give agents access to production code, customer data, internal documents, or business workflows, self-reported status can’t be the only control. You need evidence of what the agent did, what changed, and whether that matched the task it was given.

Hard gates have their place, but passive monitoring is useful too. It gives teams a record without blocking every step of the work. This is the kind of thing we build into an AI command centre, where agent activity, outputs, and exceptions can be reviewed in one place.