To monitor AI agents that have become uncontrollable, one response is required: other AIs

Agents act faster than humans can follow. A wave of startups propose to add supervisory models, an approach that is not unanimous.

The observation that crosses the industry is simple: an AI agent performs actions at a rate that no one can verify in real time. The incident at Hugging Face, where nearly 12,000 OpenAI agents coordinated faster than the teams could keep up, served as an eye-opener.

The response that emerges is to have the models monitored by other models. Apollo Research has been running Watcher since February 2026; Goodfire offers Silico, which inspects internal activations rather than outputs; Embroidery relies on the reasoning written by the agent himself. Alongside them, a whole section of application observability — Braintrust, LangChain, Judgment Labs, Arize, Galileo — is reorienting itself towards agents. Y Combinator alone has financed 106 companies in this niche.

The approach has its opponents. Simon Willison advocates for detailed, readable and verifiable execution logs, rather than an opaque model responsible for monitoring another. Avery Pennarun, from Tailscale, reminds us that the control mechanisms applicable to agents are essentially those already used to regulate human access.