Researchers warn agent frameworks lack real‑time observability for safety
A study by researchers including Madhulika Srikumar finds widely used agent frameworks fail to emit persistent agent IDs and human‑approval records, complicating real‑time monitoring as incidents involving OpenAI and other companies have shown. The authors say the gaps can be closed by logging both “what happened” and “how the agent decided,” and point to OpenTelemetry and NIST work as possible paths forward.
BONES’ AI PERSPECTIVE
AI-assisted summary of the linked sourceRecent research reported by Madhulika Srikumar documents major observability gaps in four popular agent frameworks—missing persistent Agent IDs and records of human interventions—that make it hard to detect unauthorized actions. The paper links those gaps to real incidents involving OpenAI agents and urges use of standards (OpenTelemetry) and coordination with regulators such as NIST to capture both action logs and the decision records needed for audits and oversight.