正解が一意に定まらないAIエージェントの挙動を検証する手法 Validating agentic behavior when “correct” isn’t deterministic

GitHub's engineering blog tackles a fundamental challenge in shipping agentic AI: when the same input can produce different valid outputs, traditional unit testing and deterministic assertions break down. The post outlines how teams should rethink validation when correctness is probabilistic rather than binary.

The article walks through several complementary techniques. LLM-as-a-judge uses one model to score the outputs of another against rubrics, providing a scalable proxy for human review. Scenario-based evaluation defines representative user journeys and observes the agent's behavior across many runs, treating quality as a statistical distribution rather than a pass/fail signal. Trace analysis examines the sequence of tool calls, intermediate states, and reasoning steps an agent takes, which is often more diagnostic than judging final outputs alone. Together, these approaches let teams reason about reliability even when no single ground-truth answer exists.

The context matters: as GitHub Copilot and similar coding agents move from autocomplete-style suggestions to multi-step autonomous workflows, organizations need to embed probabilistic components into CI/CD pipelines that were originally built around deterministic tests. A regression in an agent might mean a 5% drop in task success rate rather than a single failing assertion, which requires new tooling and new mental models for engineers.

This problem is not unique to GitHub. OpenAI's Evals framework, Anthropic's evaluation tooling, and third-party platforms such as LangSmith, Braintrust, and Humanloop all wrestle with similar concerns, and the industry appears to be converging on patterns that combine offline evaluation suites, online telemetry, and human-in-the-loop spot checks. There is no settled standard yet, and teams typically end up assembling bespoke pipelines tailored to their domain.

Several operational questions remain open. How sensitive should evaluation suites be to model upgrades, given that minor version bumps can shift output distributions in unpredictable ways? How should teams balance evaluation cost — every judge call incurs API spend — against coverage breadth? And how do you prevent the judge model itself from sharing biases with the agent under test, which can mask failure modes? GitHub's willingness to share its internal practices is useful precisely because these are the questions every team building serious agentic products is grappling with right now, and concrete examples from a large-scale deployment are still relatively rare in the public literature.