Goodfire launches "internal" monitors for AI agents that cost less than LLM oversight

Goodfire has released monitors for AI agents that promise oversight at a fraction of the usual price. Instead of a second language model reading every action the agent takes, the system watches the model's own internal signals as it works. An expensive check is called in only if the picture raises doubts. The announcement does not disclose a specific cost-reduction factor; the company's wording is "fraction of the cost".

How "inside-out" oversight works

The standard way to supervise an agent today looks like this: the agent writes code, browses sites and calls tools, while a second LLM reads the full log in parallel and delivers a verdict. The setup is straightforward, but the bill grows linearly. Every token of agent activity is paid for twice: first as work, then as review. If you have hundreds of agents running for hours, oversight becomes its own budget line.

Goodfire works on interpretability, the study of what happens inside a neural network between input and output. Here that research has been turned into a product. The monitor reads the model's internal representations while it runs and looks in them for signs of unwanted behavior. That could be deception, going beyond the scope of the task, or an attempt to bypass restrictions. There is no need to read the full text of the agent's actions. A signal that something is wrong with the model is enough.

A cascade instead of total inspection

The key idea is economic. A cheap internal monitor runs constantly and lets the bulk of the traffic through. The expensive tier, meaning a second model or a human, kicks in selectively. Many safety systems outside AI work this way: a simple sensor is always on duty, and complex analysis starts only when an alarm goes off.

This also raises the main question about the product. A cascade is only as good as its first filter. No one will check a violation that slips past it, and a monitor that is too sensitive will call for the expensive check so often that the savings disappear. The available materials do not describe the balance between false positives and misses.

What is known so far

• Approach: reading the model's internal states while the agent works, without separately reading the entire log.
• Escalation: a second check runs only on suspicious signals.
• Claimed effect: oversight costs a fraction of the "LLM watches LLM" setup.
• Not disclosed: exact savings percentages, the share of violations caught, the share of false alarms.
• Not disclosed: the list of supported models and whether access to weights is required.

The last point matters. You can look into activations only if you have low-level access to the model. With open weights or your own model, that is feasible. With a closed API, where only the response text is visible, such a monitor cannot be plugged in. For companies building agents on top of other vendors' APIs, the value of the product is not yet clear.

A practical view

For now this is a claim, not a verified result. The cost claims come from the company itself, and the announcement includes no independent benchmarks. But the direction looks reasonable: as agents become more autonomous, the question of "how much does it cost to supervise" becomes as important as "how much does it cost to run".

Teams already running agents in production should measure their oversight spend as a percentage of total task cost. If second-LLM review eats a noticeable share of the budget, internal monitors are worth testing in a pilot. Check not only the price but also behavior on your own scenarios: how many real violations get caught, and how many false alarms you have to sort through by hand. Teams on closed APIs would be wise to wait and see whether Goodfire offers a variant that does not need access to weights.