Show HN: A Lego-like composable Framework for RSI

Category: ai-ml

Tags: ai-agents, ide-optimization, meta-learning, benchmarking, typescript

Score: 7.5/10 (Innovation: 8, Technical: 7, Documentation: 8, Utility: 7)

Gear is an open-source TypeScript framework for recursive harness evolution (RSI) that optimizes AI agent performance on benchmark tasks by iteratively improving prompts, tools, and workflows. It combines an inner task-execution loop with an outer meta-agent-driven optimization loop, and reports measurable gains over baselines on AutomationBench and Terminal-Bench. The project is well-documented, composable, and targets practitioners building agentic AI systems.

Target audience: AI engineers, ML researchers, agent framework developers

Repository: https://rsi-gear.github.io/ · TypeScript · MIT · 207 stars

View on Hacker News