Key takeaways

  • Freeze tasks and scoring before running candidates.
  • Preserve failed trajectories and intervention.
  • Report uncertainty instead of ranking tiny samples.
01

Define the unit of success

Score the business or engineering outcome, not whether the agent produced plausible text. Specify acceptance tests, prohibited actions, time limits and reviewer rules before seeing results.

02

Sample the real task distribution

Include routine, ambiguous, adversarial and environment-failure cases in proportions that resemble production. Avoid building a suite entirely from clean tasks that already have obvious solutions.

03

Measure the operating system

Capture model, prompt, tools, permissions, retries, human interventions, latency, cost and downstream defects. Agent performance belongs to the whole configured system, not to a model name.

04

Respect sample size

Publish counts and distributions, not a false-precision composite. If evidence is insufficient to separate candidates, say so and collect the next most informative cases. A no-decision result is better than an unstable ranking.

Primary sources

  1. SWE-benchSWE-bench
  2. SWE-bench: Can Language Models Resolve Real-World GitHub Issues?arXiv

Limitations

Private suites improve relevance but reduce external comparability; preserve methodology and version every change.

Continue the research

Get evidence-led updates for teams making production AI decisions.