Agent harness engineering: LangChain lifted Terminal-Bench 2.0 score from 52.8% to 66.5% without changing the LLM
MaryamMiradi · x · 2026-11-21
A curated deep-dive into agent harness engineering, anchored by LangChain research: improving only the harness around the model lifted their coding agent from 52.8% to 66.5% on Terminal-Bench 2.0, moving from outside the Top 30 to the Top 5. The author maps 5 layers and 12 resources: agent loops (Claude Agent SDK, Codex SDK, Copilot SDK, Deep Agents), tool engineering (MCP, AGENTS.md), context engineering (Anthropic guides), memory & state (progress files + Git, Ralph Loop), and control & evaluation (verification checks, loop detection, Terminal-Bench). Practical takeaways: add a completion-time verification against the original task, detect repeated no-progress file edits, and use plain workflows when steps are known—reserve agent loops for discovery-driven tasks.
More from coding & agent
- Haijun Xia pushes back on 'UI is dead': chat is a UI too, and far from sufficient — sh_reya · 2026-11-21
- Vespa co-founders launch Hornet: solving retrieval for the agentic era — jobergum · 2026-11-21
- New Company Agent Bench: Gemini Neon leads at 87.4%, simple API harnesses beat products — philhchen · 2026-11-21
- Head-to-head coding test: Tencent Hy4 Flash finishes in 7:38 vs DeepSeek's 90 minutes — HeyToha · 2026-11-21
- Antigravity's new /teamwork feature lets you assign tasks to multiple agents at once — venturetwins · 2026-11-21
- Zed CEO mocks AI PR spam: opening 100 PRs to fix one bug isn't the flex you think — zeeg · 2026-11-21