Agent evaluation covers semantic quality and behavioral quality, with truthfulness grounded in data and goal achievement grounded in tools.
2
Evaluations need to cover single-step responses, multi-turn reasoning, tool use, planning, and the effects of actions in the agent's environment.
3
LLM-based evaluators create a second system to optimize, so teams need to manage the cost, latency, uncertainty, and operation of the evaluation process itself.
Summary
Ari Heljak presents a map for evaluating AI agents instead of treating evaluation as a loose collection of metrics. He divides the problem into semantic quality and behavioral quality. Semantic evaluation asks whether the agent's representations and responses relate correctly to reality. Behavioral evaluation asks whether its actions and tool use move it towards a goal. Each side includes single-step and multi-turn concerns. The map includes consistency, safety, values, policy adherence, context recall, faithfulness, reasoning, instruction following, tool selection, tool output quality, error handling, planning, and goal achievement. Heljak also covers cost, latency, step count, tracing, debugging, online testing, and offline testing. He argues that LLM-based evaluators create a second flow that needs its own optimization. He calls this work Eval Ops, because evaluations can become expensive, slow, and complex systems in their own right.
Agent evaluation has semantic and behavioral sides
Heljak divides agent evaluation into semantic and behavioral parts. Semantic evaluation examines how the agent's representations of reality relate to reality itself. Behavioral evaluation examines how actions and tool use contribute to the agent's goals and affect its environment. He presents this split as a way to organize the many aspects of evaluation before choosing individual metrics. The map treats what an agent says and what an agent does as related areas, since representing the world is also a kind of activity.
Single-turn semantic checks cover general response quality
For a single response, the map includes general virtues such as consistency and safety. It also asks whether the response aligns with the values of the organization's or user's stakeholders and whether it follows their policies. Heljak includes attention management, which covers whether the right context was retrieved and whether it was recalled comprehensively. He separates answer-question relevance from faithfulness and from factfulness. Faithfulness relates an answer to the reference data used by retrieval-augmented generation, while factfulness concerns its relationship with reality beyond that reference data.
Multi-turn evaluation follows conversation and reasoning over time
Multi-turn semantic evaluation examines how a conversation history develops. The evaluator may check consistency, adherence, and whether the agent stays on topic when that is required. The agent also needs to recognize when a user wants to change topics. Heljak adds reasoning traces to this part of the map, including Chain of Thought where its use is allowed. These traces provide another way to inspect sequential activities and the agent's representations of the world before it takes actions.
Tool use needs evaluation before and across chains of actions
Behavioral evaluation starts with individual actions. It asks whether the agent follows instructions, extracts tool characteristics correctly, selects the right tool, receives useful tool outputs, handles errors, and produces valid tool-call structures. For chains of behavior, the questions become broader: do the actions converge towards the goal, and is the agent's plan consistent and high quality according to the chosen measure? Heljak describes truthfulness as grounding representations in data, while goal achievement and utility ground the agent's activities in the external environment.
Goal achievement is the final behavioral measure, while other metrics are proxies
Heljak places goal achievement and utility at the end of the behavioral side of the map. The agent's actions should ultimately be judged by whether they achieve the intended goal in its environment. Measures such as tool selection, plan quality, and intermediate action quality help explain the process, but he describes them as proxy metrics. This distinction prevents teams from treating an easy-to-score intermediate behavior as the final outcome that matters.
Cost, latency, tracing, and test timing belong in the evaluation plan
The map also includes practical concerns that do not fit neatly into the semantic and behavioral split. Teams may want to optimize cost, latency, and the number of steps while the agent progresses towards its goal. Tracing and debugging help show where the agent went wrong. Heljak distinguishes tool-use error management from semantic errors in the agent's inference. He also separates offline testing during development from online testing during actual agent activity, describing them as distinct dimensions that would make the map more complicated if drawn directly into it.
LLM-based evaluation creates a second system to optimize
Many evaluation measurements use LLM-as-a-judge techniques. Heljak warns that optimizing only the agent's operative flow leaves out the cost, latency, and uncertainty of the evaluation flow itself. His double-tier view therefore covers both the LLM flow powering the agent and the LLM flow powering its evaluations. He calls the resulting discipline Eval Ops. In his description, Eval Ops is a special case of LLM Ops with different entities, software needs, ways of thinking, and resourcing requirements.
"We are calling this Eval Ops because this seems like a separate kind of thing that involves evaluations that themselves are so complicated, so expensive and slow that they sort of earn their own category of activities."11:43
Who should watch
You are building an agent evaluation roadmap and need a structure that covers more than response-quality scores.
Your agent uses tools or multi-step plans, and you need to evaluate whether its actions actually move towards a goal.
You rely on LLM-based evaluators and need to account for the cost, latency, and uncertainty of the evaluation system itself.
