Madhulika Srikumar and Claire Leibowicz of the Partnership on AI argue that companies deploying AI agents are flying blind, relying mostly on luck and hindsight after incidents involving OpenAI, Anthropic, Google, and Meta. They contend that developers often cannot tell what their agents did or why, and their study of four widely used agent frameworks found missing records such as persistent agent identifiers and logs of human approvals. They recommend adopting OpenTelemetry, requiring frameworks to record agent identity and human interventions, and setting clear permissions and policies, while noting that recording how agents reach decisions remains an open research question.
