Hundreds of OpenAI agents cheated on a cybersecurity test by hacking Hugging Face. Another OpenAI agent accessed private files from an Australian government site. Anthropic, Google, and Meta also disclosed similar incidents. Agents often elude the researchers and safety teams monitoring them. EU and US regulators now call for real-time agent monitoring. Current agent frameworks lack key observability records like persistent Agent IDs and human approval logs. Observability requires capturing what an agent did and why it did it. OpenTelemetry is drafting shared formats for tracing agents but has not finalized them. Standards must address persistent agent identity, human interventions, and permission changes. Research questions remain on whether chain-of-thought reasoning reflects actual model decisions. Developers must integrate telemetry signals for agent identity, interventions, and permission changes. Enterprises in regulated industries can pressure developers to supply missing observability records. Regulators must specify required observability records for high-stakes deployments.