AI Pulse · Under the Radar

A small open-source tool called ToolReplay is getting attention because it tackles a practical problem with AI agents, how do you know what an agent actually did once it starts calling tools, APIs, files, or internal systems?

It’s a dependency-free Python command-line tool that records an agent’s tool-call transcript, seals it with a hash chain, and can replay that record later. In plain terms, each action is tied to the one before it. If someone changes, removes, or inserts part of the transcript after the fact, the chain should show that the record no longer matches. That gives a technical team a way to inspect whether an agent stayed within its intended scope, rather than relying on a summary generated by the agent itself. The project reached 171 GitHub stars in its first day, according to the project repository.

For operators, this matters most when agents can do more than answer questions. If an agent can update a CRM, trigger payments, alter inventory, access customer data, or run workflows, you need an audit trail that can be checked after an incident. A transcript alone is useful. A transcript designed to expose tampering is more useful when accountability matters.

This is still a developer tool, not a complete governance platform. But it points to a basic requirement companies will increasingly need, clear records of agent actions, the permissions used, and whether the agent crossed a boundary. This is the kind of thing we build into an AI command centre, so teams can monitor agent activity without trying to reconstruct events from scattered logs.