← Back to Thinking

How to Use Benchmarks When Building Agents

A benchmark is not an exam — it is an answer bank. Read leaderboards to choose, copy scoring methods to build your own eval set, slice to diagnose

  • Agent
  • Benchmark
  • Evaluation

The problem

Almost every agent-building team has lived this loop: change a version, manually try three or five cases, it feels better, ship it. Two days later a user reports a new badcase, patch it again.

The problem is not that the changes are slow — it is that there is no coordinate system: you cannot tell whether a change made the system better or worse overall, nor which segment of the pipeline broke. It all reduces to one question: how does an agent know it did the right thing?

The right path: flip the loop

First define what “good” means (an eval set + scoring criteria), then regress every change against the eval set. The smallest unit of change is no longer “changed a feature” but “eval score 61 → 68, zero regressions”.

Three things to take from public benchmarks

  1. Read leaderboards to choose: compare models and approaches
  2. Copy scoring methods: borrow public benchmark criteria to build your own eval set
  3. Slice to diagnose: when something breaks, decompose a single run into nodes and localize the fault

The 7 nodes of a single execution

Any agent system — regardless of framework or model — can decompose a single request into 7 nodes: input understanding & clarification, retrieval/tool calls, context assembly, reasoning & planning, execution, result verification, output. Each node has typical failure modes — asking nothing when it should, charging ahead with ambiguous requirements, irrelevant retrieval, missing verification that lets hallucinations through.

Lessons from practice

  • An eval set is never built once: every badcase you find gets distilled into it
  • Scoring criteria must be executable: only when you can say what “correct” means do changes have direction
  • Iteration without a coordinate system only looks like progress