Open-source Gear framework automates agent self-improvement, cutting task cost by 95%
Pale-Training-6944 · reddit · 2026-11-01
The team open-sourced Gear, a meta-learning-inspired "scikit-learn for RSI" framework tackling the three recurring problems of recursive self-improvement: sampling promising changes, evaluating whether a variant actually improves the agent, and preserving diversity to avoid premature convergence. A task agent produces execution traces; a meta-agent uses them plus evaluation feedback to improve prompts, tools, workflows, and context management, decoupling search decisions from execution/evaluation so approaches like GEPA, RRSI, and Harness² become easier to implement and compare.
Results with GPT 5.6 Luna at medium reasoning effort:
- Task pass rates: AutomationBench Marketing 27.0%→40.0%, Support 17.0%→25.0%, Terminal-Bench 2.1 52.8%→62.9%
- At max effort, the optimized harness hit 88.88% objective completion vs 84.08% for Astra + Codex
- On the same optimized harness, Luna averaged $0.0568 per task vs $1.1858 for Astra — 95.21% cheaper (excluding optimization/infrastructure)
Code and paper are public; feedback welcome on candidate evaluation and diversity preservation under limited budgets.
Related event: Open-Source GEAR Framework Automates Recursive Self-Improvement for Agents(2 posts)→
More from coding & agent
- XState creator shares prompt: ask AI to find all implicit state machines in your app — DavidKPiano · 2026-11-01
- Generative coding with Grok Bot: writing the full test loop into your prompt — NickPassig · 2026-11-01
- Nous Research's Hermes Agent adds Beam data-movement integration, powered by Bittensor — markjeffrey · 2026-11-01
- Model-refusal failover chains quietly lower your real safety floor — petrusenko_max · 2026-11-01
- Curated list of 12 must-read system design and architecture newsletters — bibryam · 2026-11-01
- Context.dev raises $9M as agent web access becomes its own infra category — miky_shil · 2026-11-01