Goodfire has a new way to watch AI agents, and it does not involve hiring a second AI to babysit the first one.
The company says its monitors look inside a model while it works, rather than reading everything the agent produces after the fact. Backup review only kicks in when the internal signals look off. Goodfire is pitching this as a cheaper alternative to the standard approach, where a separate model reads an agent's entire output stream looking for problems.
The pitch matters because agent monitoring has quietly become a tax on every AI deployment serious about safety. Watching an agent's raw output after the fact means paying for a second full read of everything the first model generates. An approach that checks the model's internals instead, and only escalates when something looks wrong, promises to cut that bill without cutting corners on oversight.
Whether internal monitoring actually catches what output-based monitoring catches is the open question, and Goodfire's launch announcement does not settle it.