Study finds gaps in safety records of 4 agent frameworks

Researchers from universities, civil society and industry looked at the 4 most used agent frameworks. A framework is the software that operates AI agents. The team looked at the records that each one makes. It found large gaps. The frameworks have no agent IDs, which are names that identify each agent and stay the same. They also have no records of which humans approved agent tasks.

How it works

Monitoring of agents must have observability. This is the capacity to record the tasks of a system and turn them into signals for humans. The authors want 2 types of record. The first names the agent and records human steps and changes to permissions or memory. The next shows the reasoning of the agent and its confidence.

The proof

The study included 4 frameworks. The authors tell of gaps that make it hard to find policy violations, tasks with no approval and conditions where human oversight does not work. The source gives no framework names and no scores.

What is not known

OpenTelemetry is an open standard for software records. Its working group makes drafts of formats for agents. The authors tell the group to add agent identity that stays the same and human steps, before many companies use agents. They tell of a privacy risk when an agent ID connects to a person.

Insight: Monitor tools are only as good as the framework records

Microsoft will add agents that monitor other agents. Action State makes open tools that record work between agents. This study finds that the frameworks do not always make the records. Monitor tools must have records from the framework, and the open standard group reads as the group that selects the records they can see. Find which records a tool reads.