← Back to Harnesses
SWE-agentNotable
mini-SWE-agent
Released Jul 21, 2025
A minimal open-source coding agent from the SWE-agent team: a roughly 100-line agent class whose only tool is bash, used widely as a benchmark baseline.
- Interface
- CLI
- Pricing
- Bring your own key
- Models
- Any model
- Best Terminal-Bench
- 19.09% (v4.0) · Gemini 3.8 Flash
- Open source
- Yes
- Licence
- MIT License
Terminal-Benchv4.0 · per harness + model pairing
| Gemini 3.8 Flash | 19.09% |
| Gemini 3.7 Flash | 11.21% |
Strengths & weaknesses
Strengths
- Works with any model via LiteLLM, OpenRouter, Portkey and more
- Tiny, readable codebase with a linear message history
- Runs locally or in Docker/Podman, Singularity, bubblewrap and other sandboxes
Weaknesses
- Bash is its only tool; no IDE integration
- Best Terminal-Bench 4.0 result is 19.1% (Gemini 3.8 Flash)
Notable: appears on the official Terminal-Bench leaderboard (tbench.ai, Terminal-Bench 4.0 — the only version tbench.ai currently shows), best rank #29; only five distinct agents are listed there, so every one is top-10. Scores are TB 4.0 resolution rates (best per model across reasoning efforts), fetched 2026-10-14 — not comparable with older Terminal-Bench 2.x figures. releaseDate is the earliest GitHub release (v1.0.0.dev4); the repo was created 2025-06-28. Now at v2.