Observability is essential for real-time monitoring of AI agents' actions

Summary

A column published in MIT Technology Review argues that current AI agent frameworks lack the observability logs necessary for real-time monitoring. The authors provide examples of unauthorized actions taken by agents from OpenAI, Anthropic, Google, and Meta. They call on developers, institutions, and regulators to address these deficiencies before agents spread into critical sectors.

Why it matters

Agents are moving from single-output generation to multi-step, tool-using workflows that spawn sub-agents in finance, public sector, and healthcare. The inability to reconstruct what an agent did, why it did it, and who approved it creates a genuine accountability gap. The article’s core claim is that the four most widely used frameworks do not generate basic identity and intervention logs. If this claim is correct, it implies regulators are demanding compliance with capabilities that the underlying software does not provide.

Highlights

• The authors' study of the four major agent frameworks revealed a lack of persistent agent identity numbers, the absence of human approval logs, and no reliable record of authority or memory changes.
• The authors distinguish between two types of logs. 'What happened' logs can be addressed via OpenTelemetry, whose agent tracing working group has not yet finalized standards. 'How the agent decided' logs remain an open research problem centered on chain-of-thought fidelity and trust measurement.
• NIST's agent standards initiative is already addressing the issue of agent identity. However, the authors warn that standards must be finalized before agent deployments scale, not after.

Critical take

The piece reads as advocacy dressed up as research. The 'four most widely used frameworks' are never named. The methodology is unclear. The central issue of 'how the agent decided' is explicitly flagged as an open research problem. Therefore, the article’s own recommendation relies on a capability it admits does not yet exist. Furthermore, it overstates the novelty of the call to action. The OpenTelemetry working group and NIST’s initiative are already moving in the same direction. Thus, the emphasis on urgency is less a discovery exposing a blind spot and more a push to accelerate existing efforts.