A versioned 20-task benchmark of multi-hour software and product engineering work, designed around long-horizon implementation rather than isolated bug fixes.
Exercises sustained planning and implementation across substantial tasks. Its versioned v1.1 task package reduces ambiguity about which 20-task benchmark was run.
Result ingestion is blocked until a stable anonymous result artifact preserves exact model, agent, harness, attempt, budget, task-subset, and run identities. Client-bundle data and credential-gated logs are not a durable source contract.
Higher is better. No AgMoDB score is currently ingested; future results must remain bound to v1.1 and preserve partial, error, refusal, attempt, and budget fields separately.
Result ingestion is blocked until a stable anonymous result artifact preserves exact model, agent, harness, attempt, budget, task-subset, and run identities.
No scores available for this benchmark yet.