AI agents evade monitors, Partnership on AI finds observability gaps

The authors ground their argument in a striking analogy: a bank would never hire a thousand employees whose ID badges change daily, whose manager approvals go unrecorded, and whose changes to approval processes go unlogged. That, they say, is how AI agents are being deployed today across finance, government and medicine. Agents inherit the unpredictability of the large language models (AI systems that generate text) they are built on, but they differ in that they plan, pick tools, spawn sub-agents and take actions over longer tasks, so no pre-launch testing can anticipate every path.

The proposed fix splits into two record types. 'What happened' records are largely a software plumbing problem the authors say can be solved through OpenTelemetry, an open observability standard, though its working group has yet to finalize shared tracing formats. 'Decision' records are harder and raise open research questions, including whether chain of thought—the step-by-step reasoning a model displays—actually reflects how it decided, and whether confidence can be measured reliably. Until standards and telemetry land, the authors argue, every task handed to an agent hands over human agency too.