Show HN: AutoLearn Skills for self-improving agents (www.autolearn.dev)

AutoLearn is a system that watches an AI agent’s step‑by‑step reasoning and “crystallizes” recurring patterns into deterministic, reusable skills (i.e., code) that each agent keeps in a private skill library. No manual skill engineering or retraining is required: as agents encounter edge cases and errors, AutoLearn extracts the successful reasoning patterns and replaces repeated, expensive LLM inference with reproducible code paths. The pitch highlights a practical failure mode — roughly 10% failure per tool call — which compounds across multi‑step workflows, and argues that converting inconsistent LLM reasoning into deterministic code both improves reliability and cuts inference cost. For the AI/ML community this matters because it blurs the line between brittle workflow automation (RPA) and adaptive agents: skills are shareable across an agent fleet, continuously improve from new cases, and let agents autonomously correct and extend their behavior. Key technical implications are lower inference spend, higher end‑to‑end accuracy for multi‑step processes, and faster deterministic execution of previously heuristic LLM decisions. The approach contrasts with simple tool‑calling MCP servers by enabling genuine, evolving procedural knowledge. Practical considerations include validating crystallized code for correctness and governance, but the model promises a scalable path to more reliable, self‑improving agent automation.