Skip to lesson
Exit
Deep Dive: Graders, Rewards & Evals1 / 3

1 min lesson

Eval design for coding agents

Put this idea into your own words: "An eval you can't trust is worse than no eval, because it makes you confidently wrong."

Step 1 of 3

An eval you can't trust is worse than no eval, because it makes you confidently wrong. The whole game is building measurements whose movement you'd bet a model release on.

Start by knowing the landscape so you can speak to it. Public benchmarks tell the interviewer you read the field; internal evals tell them you understand why public ones aren't enough.

Learn more

Advanced table

Public benchmarks are a sanity floor, not the target

Eval
SWE-bench / SWE-bench Verified
What it measures
End-to-end fix of real GitHub issues, scored by hidden tests
Limit
Public, so contaminated; one distribution; saturating
Eval
HumanEval / MBPP
What it measures
Function-level code generation from a docstring
Limit
Tiny, fully leaked, not agent-assisted, far from real work
Eval
Internal task evals
What it measures
Agent on tasks drawn from real developer requests
Limit
Private and fresh, but expensive to build and label
Eval
Online product metrics
What it measures
Acceptance rate, edit-retention, follow-up rate in the live product
Limit
The truth, but slow, confounded and lagging

Public benchmarks are a sanity floor, not the target. The evals that decide a Cursor release are private, distribution-matched and tied to shipped behavior.

Learn more

Optional practice

Practice: Eval design for coding agents