Type: url
Original: https://blog.langchain.com/improving-deep-agents-with-harness-engineering/
Status: generated
Summary
A February 2026 LangChain blog post describing how the company moved its coding agent, deepagents-cli, from just outside the Top 30 to the Top 5 on Terminal Bench 2.0 — from 52.8% to 66.5% — by changing only the harness while keeping the model, gpt-5.2-codex, fixed. It describes an automated trace-analysis loop packaged as an Agent Skill, and the harness changes that helped: system-prompt guidance and a pre-completion checklist middleware that push the agent into a build-and-verify loop, injected context about the environment and time budget, a loop-detection middleware against “doom loops”, and a “reasoning sandwich” that spends more reasoning on planning and verification than on implementation. It closes with practical takeaways, including tailoring a harness to each model.