Agent decision logs are records of an AI agent’s actions and reasoning process during a session. Depending on the system, they may include prompts, plans, commands, tool outputs, intermediate choices, and diagnostic messages. They can help developers understand, troubleshoot, and audit an agent’s behavior.
These records can also contain sensitive operational context, such as details about a user’s task, development workflow, or system environment. In attacks on a service provider or its infrastructure, logs may be exposed alongside other data; their presence does not by itself prove they were copied or misused. Defenders should limit what agents record, restrict and monitor access, set retention periods, and review logs for secrets before storing or sharing them.