What problem does it solve? Multi-agent systems incur coordination overhead, context fragmentation, and brittle phase ordering, but blindly distilling them into single-agent skills can either over-constrain or under-guide the agent. This Skill analyzes the target task's evaluation metric to determine how much structure the converted skill should retain. ## Core Features & Use Cases - Metric Freedom Analysis: Scores each evaluation metric across five dimensions (equivalence strictness, output constraints, solution density, perturbation sensitivity, knowledge-to-metric distance) to produce a freedom score F. - Component Extraction: Inventories MAS agents, tools, domain knowledge, and communication patterns, then classifies each component as keep, convert, or discard based on F. - Adaptive Skill Generation: Produces a complete skill package with SKILL.md, Python tools, and evaluation test cases, with structure and length scaled to the F estimate. - Use Case: Given a multi-agent Text-to-SQL pipeline with generator, critic, and voter agents, convert it into a single-agent skill that keeps the SQL execution tools and dialect knowledge while discarding the debate and voting mechanisms. ## Quick Start Invoke /adaptive_mas_converter with the path to your multi-agent system source code and its benchmark evaluation code to generate an adaptive single-agent skill.
