We have noticed a counterintuitive fact: traditional software can be auto-tested with a single line of assert code, but Agents (AI programs that autonomously invoke tools to complete tasks) can't — the industry has split evaluation into at least 4 dimensions (correctness, relevance, tool-call correctness, safety) just to barely contain the problem.
What this is
Traditional software follows a plain logic: input "2+2", output must be "4", write an assert and run it once, done. Agents don't follow this script — when you ask "how's the weather in Beijing", the Agent may answer in one breath, or call tools first before composing the language. All three outputs are correct, but they look completely different.
Three properties make Agent evaluation hard: non-determinism (each answer may differ), no standard answer (only "good" and "not so good"), and multi-dimensionality (one answer may have correct content but messy format). Even trickier is the "right answer, wrong path" problem: when an Agent makes a wrong intermediate step but arrives at a correct final answer, traditional tests can't catch this hidden flaw.
The industry has now converged on four dimensions: correctness, relevance, tool-call correctness, and safety. LangChain's LangSmith is currently the mainstream automated evaluation platform.
Industry view
The mainstream view holds this is the inevitable direction. As models themselves approach a ceiling, "not knowing if it got it right" naturally upgrades to the next gate. LangChain launching LangSmith and major model companies doubling down on evaluation toolchains are footnotes to this judgment.
But dissent exists. One pointed critique goes straight at the LLM-as-Judge paradox: if the judge model is weaker than the tested model, the scores are unreliable; if it's just as strong, evaluation costs double. Others warn that the Agent-specific "right answer, wrong path" problem — wrong intermediate steps but a coincidentally correct final answer — is invisible to traditional assertions, producing beautiful evaluation scores and terrible real-world results.
An even more practical risk: evaluation is becoming new marketing copy. Products on the market have already started touting "99% accuracy", but evaluation standards aren't unified — this 99% and that 99% may not refer to the same thing at all.
Impact on regular people
For enterprise IT: when procuring Agents, you can no longer just watch the demo — you have to start asking "how do you test its accuracy", and evaluation reports will enter procurement checklists right alongside SLAs.
For individual careers: the capability model for knowledge workers is quietly adding one more item — not "can you use AI", but "can you judge the quality of AI output".
For the consumer market: ordinary users will increasingly see AI products flaunting "X% accuracy", but the padding behind these numbers may be even worse than phone benchmark scores.