Sign InOpen Brain
AI EngineerVideoSource Linked

The Future of Evals: From LLM as a Judge to Agent as a Judge — Aparna Dhinakaran, Arize AI

Fixed-rubric judges miss failures that emerge across long, variable agent trajectories. Arize argues for adding agent-based analysis while retaining deterministic and LLM-judge evals.

AI Engineer · Jul 24, 2026
Open Source Open MarkdownOpen JSON
Source Summary

Arize says it processes **over 100 million evals per month**; the average team runs about **12 eval jobs**, while top teams use more than 3,800 evaluators. As agents gained tools, memory, subagents, and long-running loops, their failures became harder to capture with fixed scores.

Practical Implication

Keep deterministic checks and LLM judges for stable criteria, but add trajectory-aware analysis when the path itself matters. An evaluator agent can inspect traces for repeated tool calls, loops, forgotten context, incomplete work, or inefficient execution.

Agent-Ready Context
Arize says it processes **over 100 million evals per month**; the average team runs about **12 eval jobs**, while top teams use more than 3,800 evaluators. As agents gained tools, memory, subagents, and long-running loops, their failures became harder to capture with fixed scores.

Keep deterministic checks and LLM judges for stable criteria, but add trajectory-aware analysis when the path itself matters. An evaluator agent can inspect traces for repeated tool calls, loops, forgotten context, incomplete work, or inefficient execution.

Agent-as-judge is presented as an additional eval type, not a replacement. The talk introduces Arize Signal as the implementation, but provides no comparative accuracy, cost, latency, or calibration results for deciding when it beats a fixed judge.
Context Map
benchmarkcoding#agent-evals#agent-reliability#multi-agent
Uncertainty
Agent-as-judge is presented as an additional eval type, not a replacement. The talk introduces Arize Signal as the implementation, but provides no comparative accuracy, cost, latency, or calibration results for deciding when it beats a fixed judge.