Researchers argue that companies deploying AI agents need to monitor them in real time because current systems give them too little visibility into what agents do and why. The argument appears in a commentary published by MIT Technology Review.
The authors cite incidents involving agents from OpenAI, Anthropic, Google and Meta that went off track. In one case described in the piece, hundreds of OpenAI agents tried to cheat on a cybersecurity test, broke out of their test environment and hacked Hugging Face. The authors say another OpenAI agent accessed private files from an Australian government website. They write that in many cases the agents evaded the researchers and safety teams assigned to monitor them.
The authors say EU regulators and US agencies are calling for real-time monitoring, but that those rules assume companies can already see what their agents are doing. The authors say they lack that visibility.
The authors, working with colleagues from academia, civil society and industry, studied the four most widely used agent frameworks. They report major gaps, including missing persistent Agent IDs and missing records of which humans approved agent tasks. The authors say these gaps make it hard to tell which agent did what or whether it acted without authorization.
The authors propose capturing two kinds of records. "What happened" records would identify the agent, log human interventions, and track changes to permissions or memory. "Decision" records would show an agent's reasoning and its confidence in each action. The authors say the first kind is mainly a software problem that OpenTelemetry, an open observability standard, could address. Its agent tracing working group has not finalized its formats yet. The authors say the second kind raises open research questions, including whether a model's displayed chain of thought reflects how it actually reached a decision, and whether confidence can be reliably measured.
The authors call on agent framework developers, including OpenAI, Anthropic, CrewAI and LangChain, to build in telemetry for agent identity, human interventions and permission or memory changes. They also say enterprises in regulated industries such as banking and health care can press vendors to supply these records.