Terminal-Bench 2.0: CLI & Shell Agent Leaderboard (2026)

1. The Command-Line Inflection: Why Terminal Agents Break LLM Benchmarks

The rise of autonomous developer agents—such as OpenHands, Claude Code, Cline, and Aider—has exposed an undeniable disconnect in AI evaluation: high scores on static coding benchmarks do not correlate with terminal competence. An LLM that scores 90% on HumanEval or accurately solves leetcode puzzles often collapses into recursive failure loops the moment it is placed in front of an actual bash prompt.

Terminal navigation requires an entirely different cognitive profile:

• Non-Stationary Environmental Feedback: Every command mutates the underlying OS state. Files are created, permissions shift, environment variables get overwritten, and ports open or close.
• Exit-Code Sensitivity: A silent return code of does not mean a task succeeded; conversely, standard tools like grep or diff emit non-zero return codes under normal operational conditions.
• Partial Observability: Long compilation traces, truncated logs, and interactive pagers (such as less) demand proactive buffering and deliberate grep filters rather than raw prompt ingestion.

2. Architecture of Terminal-Bench 2.0: The 89 Sandboxed Linux Tasks

Terminal-Bench 2.0 standardizes agent evaluation across 89 meticulously designed real-world Linux engineering tasks. Each task launches inside a pristine, headless Docker container running Debian or Ubuntu with pinned system packages and specific broken system configurations.

3. Four Forensic Failure Modes: Why Frontier Models Fail in Shells

In our independent evaluations of hundreds of agent runs across Terminal-Bench 2.0, we have identified four universal failure mechanisms that differentiate true frontier performers from brittle completion models:

4. Verified October 18, 2026 Leaderboard Analysis

The table below reflects 100% verified empirical scores across the 89-task Terminal-Bench 2.0 evaluation corpus under uniform agent scaffolding constraints:

5. How to Reproduce This Evaluation Locally

All Terminal-Bench 2.0 tasks are completely reproducible on any standard x86_64 or ARM64 Linux/macOS machine equipped with Docker:

# 1. Clone the standardized Terminal-Bench evaluation repository
git clone https://github.com/laude-agents/terminal-bench-eval.git
cd terminal-bench-eval

# 2. Setup your virtual environment and install test dependencies
python3 -m venv .venv && source .venv/bin/activate
pip install -r requirements.txt

# 3. Pull the standardized Linux sandbox task images
docker pull ghcr.io/vibecoderjournal/terminal-bench-sandbox:v2.0

# 4. Run the full 89-task evaluation against your chosen model
python evaluate.py     --model "claude-3-7-sonnet"     --harness "harbor-agent"     --tasks "all"     --timeout 180     --output-dir ./results/2026-10-18/

6. Known Boundaries: What Terminal-Bench Does Not Measure

Every rigorous benchmark is intentionally narrow. Reading Terminal-Bench as an all-encompassing verdict on software development prowess is a fundamental mistake:

• Code Readability & Architecture: The harness only evaluates whether the final verification script returns exit code 0. It does not grade elegance, adherence to Clean Code standards, or software design patterns.
• Multi-Day Development Cycles: Tasks are calibrated to execute within 3 to 15 tool turns. It does not test how agents handle months of repository evolution or distributed architectural redesigns.
• Graphical UI Workflows: Terminal-Bench is strictly headless; it contains no evaluation of CSS layouts, DOM trees, or frontend interactions.