Langfuse says AI agents can be incorrect while system metrics stay green
In its example, an agent says that it made a refund but does not use the refund tool.
A Langfuse engineer says teams must test AI agents in a loop. They check records of use and get expert help. They mix AI judges with set rules. They stop a release if a test shows a problem. For teams that ship agents, tests must check the work of the agent.
Sources
Posted
