tbench

What problem does it solve?

This skill enables automated benchmarking and failure analysis for Unix-based agents using Terminal-Bench 2.0, providing a repeatable workflow to measure performance and identify bottlenecks.

Core Features & Use Cases

• Unified benchmarking: Run full or partial Terminal-Bench benchmarks against Unix agents locally (Docker) or in Daytona cloud sandboxes.
• Configurable workloads: Control datasets, concurrency, timeouts, and agent settings via environment variables (TB_DATASET, TB_CONCURRENCY, TB_TIMEOUT, TB_ENV, TB_TASK_NAMES, TB_ARGS).
• Result tracking: Automatically collect run results, logs, and per-task outcomes for performance analysis and leaderboard submissions.
• Use Case: A team validating agent robustness across tasks like chess moves or maze navigation by executing scripted benchmark runs and comparing results.

Quick Start

Use the tbench skill to run the full terminal benchmark suite or targeted tasks with Makefile targets and TB_ARGS.