A Correct Answer Can Still Invalidate Your AI Agent Eval, Microsoft Dev Blog Warns
WirelessLife · x · 2026-09-18
A Microsoft developer blog post argues a passing agent evaluation can prove the wrong thing: the answer may be correct while the measurement is invalid. Agents can answer from internal knowledge — or retrieve evidence from prompts, local source, caches, tools, or services, and each route demonstrates a different capability that the final answer doesn't reveal.
Key points:
- Define the measurement first: before restricting tools, decide what the eval should prove — researching an unfamiliar API, using repo files, or testing existing model knowledge.
- The sandbox must block the knowledge under test: if you're measuring what the model knows, retrieval must be prevented; otherwise the score reflects everything available during the run.
- Inspect surprising passes: an impressive-looking pass may just be retrieval, not knowledge.
Bottom line: your eval is only as good as its sandbox, or you're answering a different question than you think.
More from coding & agent
- Claude Code 2.1.275 release imminent — ClaudeCodeLog · 2026-09-18
- Legora's take: there is no best model — lawyers write evals and Legora BAR picks the winner — soleio · 2026-09-18
- Research with Agents: reflections and lessons from using AI agents in scientific work — _xiang_chen_ · 2026-09-18
- Addy Osmani: Running Agents in Old Codebases — Brownfield Agentic Engineering — rseroter · 2026-09-18
- Google's Stellar Colosseum: many-agent harness proves new math theorems — IgorCarron · 2026-09-18
- Official MCP extension lets agents discover and load Skills directly from MCP servers — solyarisoftware · 2026-09-18