Software World Model: leaderboard

Agents perform tasks in application databases. Outcomes distinguish completing the requested change from damaging unrelated state, and retain per-application and per-skill denominators.

Source: raw.githubusercontent.com.

Benchmark results

Results are published separately for each benchmark or evaluation setting.

Interactive version: theaggregate.ai/benchmark?slug=software-world-model · How It Works · Data refreshed daily, snapshot 2026-11-17.