Harness tuning lifts Qwen3.6-27B legal agent pass rate from 67% to 85%, study finds
sarahookr · x · 2026-09-09
Research shared by sarahookr argues agent performance is wrongly attributed mostly to the underlying model: the harness can only recover what a model can already do, not create capability it lacks.
Key numbers: on the Harvey Legal Agent Benchmark, harness optimization alone moved Qwen3.6-27B's pass rate from 67% to 85%, with post-training pushing it only to 88%.
The author's framing: optimize the entire stack, like a chef wanting both the best ingredients (model) and the best oven (harness), rather than fixating on one.
Related event: Harness optimization lifts legal agent benchmark from 67% to 86%(3 posts)→
More from coding & agent
- Dev ships a delightful free shadcn UI for agent run and tool-call traces — round · 2026-10-27
- Musk hails Grok Bot as 'incredibly competent chief of staff' as user automates inbox triage — elonmusk · 2026-10-27
- shadcn's /improve skill paired with Grok on SuperGrok Heavy for plowing through issues — shadcn · 2026-10-27
- Paradigm Launches Envoys: Your Claude Code and Codex on Your Research Graph, Free — tensorqt · 2026-10-27
- AXL launches remote MCP server exposing 1,284 sales and marketing APIs to Claude and Codex — EdTech_Padre · 2026-10-27
- GitHub's Thomas Dohmke ships Entire, a Git forge built for agent-human collaboration — carrycooldude · 2026-10-27