跪拜 Guibai
← All articles
Artificial Intelligence

Agent Evaluation Is Moving From Grading Answers to Auditing Behavior

By 美团技术团队 ·
Read original on juejin.cn ↗ Google Translate ↗ Alt translation

As Agent frameworks move from single-turn Q&A to multi-step task execution, evaluation that only checks final answers misranks unreliable, costly Agents as equal to stable ones. Teams building on Claude Code, OpenClaw, or internal Skill platforms need trace-level behavioral evaluation and binary-rubric alignment to turn subjective judgment into repeatable, automated quality gates.

Summary

Agent evaluation has outgrown leaderboard scores and demo impressions. Meituan's Turing team frames it as a precision instrument that must answer not just whether an Agent succeeded, but how — examining the planning path, tool-call stability, token cost, and safety risks in every run. The core formula is observation plus evaluation equals continuous iteration; without full-trace logging, most failures are invisible and therefore unfixable.

The team's central methodology is bridging, not piling up metrics. Business KPIs like DAU or retention cannot map directly to model capability scores, so they insert a task-system bridge layer. Subjective evaluation is aligned through a "dictator" model — one strong role forces product, ops, and engineering to converge on the same binary rubrics (yes/no/unknown), which Meituan used to push human-machine agreement from 62% to 92% on one project and to 99% on another.

Long-horizon Agents like Claude Code and OpenClaw are forcing the biggest shift. Evaluation is no longer about whether the reply sounds good; it is about whether the task got done and whether the execution trace is repeatable. The audience for evaluation tools is also expanding from a small circle of ML engineers to anyone who writes a Skill, which demands simpler, automated, pipeline-integrated evaluation infrastructure with full-link replay, sandboxed execution, regression gates, and AI grading engines that amplify a core evaluator's judgment rather than replacing it.

Takeaways
Agent evaluation must cover four layers: result correctness, process stability, token and time efficiency, and safety/authorization risks.
Two Agents can produce the same final answer but have radically different engineering value; only process-level evaluation catches the difference.
Full-trace observation is a prerequisite — problems that leave no log are nearly impossible to diagnose or fix.
Business KPIs and model capability scores do not map directly; a bridge layer of task-system metrics is required to explain why a model upgrade did or did not move the business needle.
Subjective evaluation alignment works best with a single "dictator" who forces cross-functional agreement on binary rubrics (yes/no/unknown), then iterates until human-human and human-machine agreement exceed 85-90%.
Meituan's Digital Station Master project grew from 20 evaluation metrics to nearly 200 over a year, driven entirely by Good Case and Bad Case analysis.
In vertical domains where public high-quality data is scarce, domain-expert input is the cold-start lever; divergent expert opinions can be productized as separate Agent strategy branches.
Long-horizon Agent evaluation shifts the unit of analysis from (query, answer) to (prompt, expected_behavior, trace), making task completion and trajectory quality the primary objects.
Skill evaluation's core pain point is that most people do not know how to write a good Skill and lack lifecycle tooling; the user base for evaluation is expanding to anyone who creates or modifies a Skill.
Long-horizon evaluation infrastructure needs full-link replay, Case management, sandboxed execution, Rubric-driven AI grading, regression triggers, and pipeline gates to move beyond one-off project support.
Conclusions

Meituan's framing of evaluation as a "precision measuring instrument" rather than a leaderboard exercise reflects a broader industrial reality: Agent reliability is an engineering problem, not a model-capability problem, and the bottleneck is observability.

The "dictator" model for aligning subjective evaluation is a deliberate rejection of democratic consensus in favor of speed and consistency — a trade-off that many Western teams resist but that Meituan credits for lifting human-machine agreement from 62% to 92%.

Binary-rubric decomposition (yes/no/unknown) treats the proportion of "unknown" verdicts as a leading indicator of rubric quality, turning ambiguity into a signal for where the evaluation framework itself needs redesign.

The observation that divergent expert opinions on what constitutes "good" can be productized as separate Agent strategy branches suggests a path to handling long-tail scenarios without forcing a single gold-standard answer.

Long-horizon Agents invert the evaluation cost structure: richer execution traces make human evaluation a bottleneck, which pushes the workflow from human-led to machine-led evaluation — but only after core evaluator standards are encoded into rubrics.

The Skill evaluation gap is not primarily technical; it is a tooling and literacy gap. As Skill creation drops to near-zero-code, the population that needs to evaluate output will include product managers and operators who have never run an eval before.

Concepts & terms
Trajectory Evaluation
Evaluating an Agent by examining the full execution trace (planning steps, tool calls, intermediate states), not just the final output. Contrasts with Response Evaluation, which only judges the answer.
Rubric Binarization
Decomposing vague evaluation criteria into yes/no/unknown checks. This reduces inter-rater disagreement and enables reliable automated scoring; the proportion of 'unknown' verdicts signals where the rubric itself is poorly defined.
Long-horizon Agent
An Agent designed to complete multi-step, open-ended tasks over extended time spans, requiring dynamic planning, multiple tool calls, and intermediate-result processing. Examples include Claude Code and OpenClaw-based agents.
Task (in Agent evaluation)
A single test case with a defined prompt, expected behavior, and success criteria. Running the Agent produces a trace; comparing the trace against expected behavior yields an evaluation triplet (prompt, expected_behavior, trace).
Human-Machine Alignment
The degree to which an automated evaluator's judgments match those of a calibrated human evaluator. High alignment (e.g., >90%) is a prerequisite for replacing manual review with machine-led evaluation at scale.
Source: juejin.cn ↗ Google Translate ↗ Backup ↗