Terminus 2 (Terminal-Bench)

Terminus 2 is a minimal, terminal-native AI agent designed by the Terminal-Bench team.

Description

Unlike complex agents with multi-step reasoning engines, Terminus 2 takes a "raw" approach by giving the LLM direct access to a tmux session. The model sends commands as text and parses the terminal output itself. Despite its simplicity, it performs remarkably well on terminal-based benchmarks.

Links

• Terminal-Bench GitHub
• Reference Blog Post

Alternatives

• OpenHands
• Devin

Backlog

• Add setup instructions for the tmux-based environment.