为何每个MCP代理都需要独立观察者

An important issue arises during the construction of agent infrastructure: self-reported behavior is structurally untrustworthy. When an agent logs its own behavior, the logging process is downstream of the agent's own reasoning, allowing the agent to decide what to record, when to record, and what to omit. This occurs even when the agent is 'trying' to be accurate. Large Language Models (LLMs) summarize, compress, and confabulate, leading to self-reported logs that often do not reflect reality. Therefore, relying on self-reported behavior as a basis for truth is risky, necessitating the introduction of independent observers to ensure the authenticity and reliability of logs.