Gear evolves your agent harness instead of swapping models — up to +13 points pass rate on the same model
Pale-Training-6944 · reddit · 2026-11-05
- The author released Gear (open source, MIT): instead of swapping models, it automatically evolves the agent harness — instructions, tools, skills, workflows, context management — against a benchmark you define.
- How it works: a Meta agent reads execution trajectories plus eval feedback, proposes candidate harness changes, and Gear evaluates candidates and promotes only winners; the loop is Diagnose → Evolve → Evaluate → Promote, with per-component restrictions.
- Results (same model, GPT-5.6 Luna medium): AutomationBench Marketing 27.0%→40.0%; Support 17.0%→25.0%; Terminal-Bench 2.1 52.8%→62.9%.
- v0.2.2 lets you write diagnosis/candidate-generation logic in Python; the README includes an animated replay of optimized vs baseline harness. Repo: rsi-gear/gear.
More from coding & agent
- Gradium and SurrealDB demo a fully on-device research agent with memory and streaming TTS — mattturck · 2026-11-05
- webIQ demo shows the evidence object returned to models, stressing freshness signals — VeryWellVersed · 2026-11-05
- Paper: rewriting reasoning histories into compact records cuts agent token costs — Anton Hristov · 2026-11-05
- KernelSeek: question-driven agent hits 1.13x FlashAttention-3 where tool-rich agent stalls at 0.50x — Snowflake · 2026-11-05
- Orbit beta puts Claude Code, Codex and Grok in one app with multi-account switching — Available-Ganache-71 · 2026-11-05
- Stop saying 'write code': 8 senior-engineer prompt templates for Claude — aitrendz_xyz · 2026-11-05