GEAR 发布可自我改进的智能体 harness，公开任务完成率从 75.37% 升至 88.88%
Show HN: I made an agent can self-improving while working
GEAR 为智能体 harness 引入递归自我改进（RSI），利用任务反馈持续优化技能、工具和工作流，支持 Codex、Claude Code、DSH、Pi 等已有智能体。
Don’t let your work wait. GEAR brings recursive self-improvement (RSI) to your agent’s harness — using task feedback to continually refine skills, tools, and workflows for the work that matters.
Bring the agent you already use.
CodexClaude CodeDSHPi+
GEAR-tuned harness.
Beyond general-purpose agents.
On these two real tasks, Luna max with a GEAR-tuned harness completes requirements that Astra max with the general-purpose Codex agent misses. Watch the difference side by side.
CASE 01AutomationBench / Marketing
Landing page alerts
CASE 02AutomationBench / Marketing
Featured snippet opportunities
Small model.
Serious results.
An optimized harness can change what an agent gets done. Explore the results reported in our open experiments.
Marketing · 100 public tasks
Objective completion
Before75.37%
Luna · original DSH · medium
After GEAR88.88%
Luna · optimized DSH · max
Task pass rate also increased from 27% to 53%. Astra max with Codex scored 84.08% objective completion and 57% task pass rate.
How GEAR works.
Follow a real AutomationBench optimization: inspect failed runs, propose harness changes, evaluate candidates, and retain the versions that perform better.
Help more people discover GEAR.
来源：Hacker News · yehuitang0316.github.io