1,400-line teaching agent harness scores full marks on Terminal Bench 2.1
PawelHuryn · x · 2026-11-06
Paweł Huryn's open-source teaching harness, mini-harness (1,400 lines of Python, 2 dependencies), achieved full scores on all 89 Terminal Bench 2.1 tasks across 3 runs, beating Pi (155K context), mini-swe-agent (4K), and OpenCode 2.
Despite its tiny size, it includes goal mode, permissions, skills, MCP server support, and Anthropic's server-side compaction — built to explain how a harness works end to end. It currently speaks Anthropic's message format; the author plans to benchmark it with DeepSeek and add support for OpenAI endpoints and any model.
Related event: 1400-line Python teaching harness scores full marks on Terminal Bench 2.1(2 posts)→
More from coding & agent
- Coding agent overhead was ~90% of spend; task-based attribution cut it to 45% in 5 weeks — entelligenceai17 · 2026-11-06
- Viktor gives its AI employee an email address — just CC it to delegate work — hey_abusiddik · 2026-11-06
- New podcast episode on the decision layer for coding agents: local models routing between code and LLM reasoning — TheZachMueller · 2026-11-06
- TypeSafe engineer: stop calling Jev a decision model, wrong abstraction hurts reliability — iamrobotbear · 2026-11-06
- Latent Space #12: mitsuhiko returns to discuss state of agentic coding — mitsuhiko · 2026-11-06
- Do You Switch Agent Harnesses Per Task? Reddit Thread Probes Harness Optimization — Pale-Training-6944 · 2026-11-06