mkadri85/guardplane

Guardplane is a small code library that acts as a safety layer for teams running fleets of AI agents, the kind of automated systems that can take real actions like calling tools, editing files, or making decisions on their own. The problem it addresses is that AI agents sometimes go wrong in production, and without something watching them, a broken agent can keep taking damaging actions before a human notices. The README points to a real example from April 2026 where a coding agent deleted a company's production database and its backups in about nine seconds. The library is not a full framework you build your whole system around. Instead it is a lightweight skeleton you wrap around whatever AI agent setup you already have. It works through a few connected pieces: a signal plane that records every decision an agent makes so it can be replayed later, a reasoning layer that runs simple checks to detect whether an agent is stuck in a loop, spending too much money, producing too many errors, or quietly getting worse, a confidence gate that decides whether a known problem can be fixed automatically or needs to be handed to a human, an action layer that can reroute an agent to a different model, roll it back, or pause it, and a kill switch that can stop one agent or the entire fleet immediately. A newer feature called the burn rate breaker watches for agents whose judgment is slowly degrading rather than failing all at once. It compares each agent's current failure rate to its own recent history, and if that rate climbs too fast, the agent is switched into a mode where it can still work but its output becomes a suggestion for a human to approve rather than an action it takes on its own. To try it out, you clone the repository, install its dependencies with npm, and run the included demo, which walks through a healthy agent, a looping agent, a cost overrun, and a quality decline, showing how each situation is handled. The project has zero runtime dependencies and is written in TypeScript, and it is offered as a companion piece to a blog post the author wrote called Who Operates the Operators.