The Future of Evals: From LLM as a Judge to Agent as a Judge — Aparna Dhinakaran, Arize AI
Source
AI Engineer
Author
AI Engineer
Date
Key takeaways · AI-distilled
Arize runs over 100 million evals per month across its customers; the average team runs about 12 eval jobs, while the top teams run more than 3,800 different evaluators.
Aparna Dhinakaran argues the thing being evaluated changed underneath first-gen evals: from answering a prompt in 2023, to tool calls, reasoning, and deep research in 2024, to loops on real-world data with sub-agents on long-horizon tasks — each jump producing a fundamentally dif
Arize's own in-product AI agentAn AI system that doesn't just answer once but works toward a goal in a loop — taking actions, reading the results, and deciding what to do next.Full definition →, Alex, surfaced the failure modes classical evals miss — forgetting context windowThe maximum amount of text a model can consider at once — its working memory for the current conversation or task.Full definition →, not knowing when a task was done, getting stuck in loops — because every user interaction generates a new dynamic UI and thus a new trajectory.
Her core claim: LLMA large language model — the neural network behind tools like Claude and ChatGPT, trained on huge amounts of text to predict what comes next.Full definition →-as-a-judge applies a fixed rubric with fixed scores, which breaks when an agent takes a completely different trajectory on every run — so the answer is agent-as-a-judge, with the future being all three tiers (deterministic evals, LLM-as-a-judge, and agent-as-a
Arize released Signal, a long-running agent-as-a-judge that reads submitted traces, discovers subtle failure patterns like repeated tool calls and inefficient trajectories, and can then open a PR with a proposed fix.
Terms in this piece · Glossary
LLM-as-judge — Using one model to score another's output against a rubric, so quality can be measured at a scale human grading cannot reach.
AI agent — An AI system that doesn't just answer once but works toward a goal in a loop — taking actions, reading the results, and deciding what to do next.
context window — The maximum amount of text a model can consider at once — its working memory for the current conversation or task.
LLM — A large language model — the neural network behind tools like Claude and ChatGPT, trained on huge amounts of text to predict what comes next.
Why it matters
If you're building or grading AI agents, this lays out why static and LLM-as-judgeUsing one model to score another's output against a rubric, so quality can be measured at a scale human grading cannot reach.Full definition → evals break as agents gain reasoning and tool-use, and makes the case for agent-as-a-judge systems that surface unanticipated failure modes and can even open PRs to fix them.
Key quotes
“Evals have gone from the new skill that every PM and every AI engineer has to learn to the thing that every serious AI team is betting on.”
Aparna Dhinakaran
“Every one of these was actually a massive jump in complexity, and we didn't just make the problem harder, we actually got a fundamentally different type of problem.”
Aparna Dhinakaran
“What if the best way to an evaluate an agent was actually with an agent.”
Aparna Dhinakaran
“LLM as a judge just gives you a fixed rubric with these fixed scores.”
videoWhy AI Agents Need Million-Token Context — Thomas Wolf & Olive Song, MiniMax
videoYour company brain will leak secrets: how we stopped it for big banks — Tanmai Gopal, PromptQL
videoTethered: Our Agents Are Us — Shu Fang, Two Sigma
videoAgents' next frontier: agent-to-agent and network effects — Jean-Denis Greze, Town
