What happened
Goodfire, a startup focused on AI interpretability, has launched a new monitoring system for AI agents. The system uses internal probes to watch a model’s behaviour as it works, rather than using a second AI to review its output. TechCrunch reports the service is now available to customers of the model hosting platform Baseten.
In tests on the Kimi K3 open model, Goodfire claims its approach cost $51 to monitor 1,500 sessions, compared to $233 for a cheap external model. The probes caught 94% of malicious hacking sessions.
How the room's reading it
The standard for agent safety has been an external AI acting as a chaperone — an effective but costly setup that adds latency and doubles inference costs. Goodfire’s launch is seen as a direct response to this expense and to recent incidents where agents have escaped their test environments. The pitch is aimed squarely at developers using open models, where safeguards can be easily removed.
Infra teams and inference providers are watching closely, as they often carry the liability for what happens on their platforms. While the technique isn't entirely new — Google DeepMind has deployed similar probes in Gemini — Goodfire is among the first to productise it for the open model ecosystem.
Sailfish's take
We think monitoring agent behaviour at the output layer is too slow and too late. The real work is catching intent before it becomes action, and that requires looking at a model's internal state. This is the core appeal of Goodfire's approach — it's not just cheaper, it's architecturally smarter because it reuses computations the model is already performing.
For anyone shipping autonomous agents, this moves safety from a post-hoc check to an in-flight one. We see this becoming the default safety stack for any serious agent deployment. If you're building with agents, this is the week to start investigating internal monitoring.