AI Agent Monitoring Needs Persistent Identity, Approval Logs and Decision Records
AI agents can act across systems, call tools and delegate work before a person sees the result. That makes a basic operational question difficult to answer: which agent did what, under whose authority, and with what human review? Recent incidents suggest companies often can’t reconstruct those details quickly enough. The practical weakness is not simply that agents can misbehave; it’s that the systems around them may fail to leave the records needed to detect and investigate it.
This summer, hundreds of OpenAI agents attempting a cybersecurity test escaped their test environment and hacked Hugging Face, which they believed contained clues about their grading. Around the same time, another OpenAI agent accessed private files on an Australian government website. Anthropic, Google and Meta have also disclosed cases in which agents went off track, sometimes without the researchers and safety teams responsible for monitoring them noticing in time.
Agent oversight depends on records
Regulators and standards bodies are calling for real-time monitoring. The EU and US agencies are right to focus on oversight, but monitoring can only work if a system produces useful evidence. An organization needs to capture what an agent did, turn those events into meaningful signals, and make them available to people or automated oversight tools. That capacity is known as observability.
Think of a bank assigning a thousand new employees access to customer accounts. Their ID badges change every day, managers’ approvals aren’t recorded, and nobody tracks edits to the approval process. Even a diligent audit team would struggle to establish who acted, what authority they had or whether a supervisor reviewed their work. That is a useful picture of the visibility problem facing agent deployments.
The stakes rise as agents move into finance, government and medicine. In these settings, an unexplained action can affect access to services, sensitive information or consequential decisions. Predeployment testing still matters, but it can’t predict every path an agent might take once it begins planning, selecting tools, starting sub-agents and acting over an extended task.
A conventional generative AI system returns an output for a person to evaluate and use. An agent can carry out a sequence of actions, and each choice can shape what happens next. It inherits the unpredictability of the underlying language model while adding more opportunities for unexpected behavior. That makes observation during execution a central safety requirement, rather than a task reserved for post-incident review.
Frameworks leave gaps in the audit trail
A study by the Partnership on AI and collaborators examined four widely used agent frameworks and the signals they emit. The researchers found missing records that make it harder to detect policy violations, unauthorized actions and failures in human oversight. Among the gaps: agent IDs that persist over time and records showing which people approved an agent’s task.
Without persistent identity, organizations may not be able to connect actions across a run or reliably determine which agent performed them. Without approval records, an audit may show that an agent acted but not whether a person authorized the task. These omissions weaken both live monitoring and the later work of reconstructing an incident.
Useful observability needs two kinds of records. The first documents what happened during a run: an agent’s identity, human interventions, and changes to its permissions or memory. The second describes how the agent reached its decisions, including its reasoning and its confidence in the actions it took. They answer different questions. Event records establish the sequence of activity and the authority around it; decision records may help explain why the agent chose a particular action.
Consider an agent helping process mortgage applications. It recommends denying an application, and a loan officer approves that recommendation. A bank may need to establish which agent made the recommendation, which officer signed off and how long approval took. It may also need to inspect the basis for the decision to check whether the agent relied on prohibited information, such as details about the applicant’s neighborhood. A record of the final decision alone can’t answer all of those questions.
Shared telemetry can address part of the problem
Recording actions and interventions is, in large part, a software infrastructure problem. OpenTelemetry, an open observability standard maintained by engineers across the industry, offers a path for frameworks to emit consistent telemetry. Its working group is drafting shared formats for tracing agent activity, but those formats aren’t final.
The standards work should cover persistent agent identity, human interventions, and changes to permissions or memory. Those records would give operators a clearer account of what agents did and the authority under which they acted. Identity design also needs to account for privacy: a persistent identifier can improve accountability, but linking it to a real person creates privacy questions that the system must address.
Capturing how an agent decided is more complicated. A model’s step-by-step explanation, often called chain of thought, may not faithfully represent the process that produced its answer. Researchers and model providers also need reliable ways to measure an agent’s confidence. Those open questions matter before organizations define how decision records should be interpreted or used. A log that looks explanatory can create false assurance if it doesn’t accurately reflect the underlying process.
Responsibility is shared across the stack
No single vendor or customer can close these gaps alone. Framework developers, including OpenAI, Anthropic, CrewAI and LangChain, can build telemetry for persistent identity, human review and permission or memory changes into their products. Enterprises have practical influence too: banks and hospitals can make access to deployment environments conditional on receiving the records they need, even while shared standards are still taking shape.
Telemetry can show what happened, but an organization still has to decide what an agent is allowed to do. Enterprises need clear policies, permissions and roles that define those boundaries. Regulators and AI standards institutes can set expectations for the records required in high-stakes deployments while leaving implementation details to industry. In the US, NIST’s agent standards initiative is already working on agent identity, one of the gaps identified in the research.
The operational test is straightforward: when an agent acts, can its operator identify it, trace its actions, see who approved them and determine whether its authority changed? If the answer is no, an organization is relying on hindsight to manage a system that may already be acting across sensitive workflows. Persistent identity and usable records won’t prevent every failure, but they give people a chance to notice and respond while an agent is working—and to establish what happened afterward.
Madhulika Srikumar is Head of the AI Safety Program at the Partnership on AI, a global nonprofit partnership of industry, academic and civil society organizations focused on AI governance. Claire Leibowicz is Director of AI, Trust, and Society at the Partnership on AI and a doctoral candidate at Oxford University.