AI
Agent Evaluation Is Not Model Evaluation
Agents contain LLM calls, tool calls, and retrieval steps that unit tests cannot reach. Eval-powered tests fill that gap, the same way test-driven development works for deterministic software. Without them, teams have no way to detect bad outputs before they hit production. In non-deterministic systems, the failure mode is not a crash; it is an agent silently doing the wrong thing for days.
Anatomy of a Multi-Agent System
The travel agent example runs on Azure AI Foundry with two components: a planner agent that retrieves user preferences and calls tools for flights, hotels, and activities, and a reservation agent that calls Expedia, Yelp, and OpenTable. They hand off rather than share credentials. Authentication across agents is unsolved, so the planner never holds the user’s Expedia login. The multi-turn travel agent live demo (07:21) shows per-step metrics alongside each conversation turn.
The Eval Metric Taxonomy
Single-step metrics cover hallucination, completeness, tool selection, and context attribution (a RAG-based metric). Agents have been observed calling the same tool twice instead of two different tools; tool selection errors like that must be caught before production. Trajectory metrics go further: action advancement tracks whether the agent is progressing toward the goal or looping on the same questions, turn frequency flags a planner taking twenty turns when two would do, and topic deviation catches an agent volunteering unsolicited recommendations. “There’s no silver bullet here,” Yashad said. Every metric must be adapted to the specific use case.
CI/CD, Real-Time Guardrails, and Distributed Tracing
Eval metrics sitting outside CI/CD are fragile. Teams have shipped agentic features, then rolled them back a month later because a new change regressed earlier instructions. Galileo built lightweight evaluation models that take any LLM-as-judge metric and run it as a guardrail in under 200 milliseconds, making real-time enforcement practical even as agents start modifying databases and ticketing systems. For multi-agent flows, the platform adds distributed tracing: every agent’s steps are collected into a single graph, and the system flags where handoffs went wrong.
Notable Quotes
One agent to rule them all is not the reality. Yashad · ▶ 2:27
There’s no silver bullet here Yashad · ▶ 11:24
eval powered tests are critical Yashad · ▶ 1:36
Key Takeaways
- Eval-powered tests catch LLM call failures that unit tests cannot reach in non-deterministic agent code.
- Multi-turn agents need trajectory metrics like action advancement and turn frequency, not just output quality scores.
- Galileo’s sub-200ms guardrail models convert any LLM-as-judge metric into a real-time production circuit breaker.