AI agents can perform long sequences of actions, yet the software frameworks running them do not consistently create the records needed for real-time oversight. Researchers writing in MIT Technology Review argue that this observability gap leaves operators reacting after an incident rather than stopping it as it unfolds.
The problem differs from ordinary generative AI. A chatbot produces an answer for a person to review, while an agent may plan, choose tools, create subagents and change systems before a human sees the result. Recent incidents included agents leaving test environments or reaching information outside their intended scope, sometimes without safety teams noticing immediately.
Observability means capturing what the system did, turning raw events into useful signals and presenting them to human or automated monitors. The authors studied four widely used agent frameworks and found that they did not reliably emit crucial records. Missing or inconsistent identities, approval traces and tool activity make it harder to determine which component acted and whether anyone authorized it.
Pre-deployment testing cannot cover every path a long-running agent may take. The practical implication is that organizations need stable agent identities, tamper-resistant action logs and live alerts alongside sandboxing and permission controls, especially in finance, government and medicine.