learn-skill

Build a reusable skill package by capturing live web evidence into tiered bundles and distilling it into rules, then validating execution through isolated agents to prevent hallucination. The orchestrator enforces file-only handoffs for strict agent isolation.

Closed-book skill evaluation runs executor agents without access to reference corpus, memory, or judge gold answers, preventing hallucination by requiring independent judging. Judges score outputs and provide repair targets, while repair agents revise using only allowed failure summaries.

Set up isolated subagents by configuring a main orchestrator that coordinates while subagents communicate strictly via persisted files for skill orchestration. This enforces whitelisted inputs for executor, judge, and repair agents while maintaining evidence-tier separation.

Persistent memory tracks iterative revisions across train, dev, canary_hidden, and final_hidden evaluation loops to evolve domain skills. The final_hidden gate enforces non-leaking constraints to ensure the distilled skill package passes strict validation.

Closed-book skill distillation is limited to bounded professional workflows that can be distilled into a reusable skill package and validated via evaluation loops. It requires strict agent isolation and non-leaking final_hidden constraints, making it unsuitable for open-ended or unbounded tasks.