How bingreeky/maas works
Project type and purpose
• MaAS is a research‐oriented Python library and CLI tool for “Multi-agent Architecture Search via Agentic Supernet.”
• It dynamically samples and optimizes multi-agent system architectures for diverse user queries, leveraging large language models (LLMs) and feedback.
• It includes experiment harnesses for benchmarks (GSM8K, HumanEval, MATH) and integrates with multiple LLM providers, retrieval backends, and tool plugins.
File structure
• Top-level directories:
• maas/ ← core library
• examples/maas/ ← experiment entrypoints (optimize.py)
• config/ & .maas/ ← configuration files (YAML)
• assets/ ← figures for README
• maas/ext/maas/ ← benchmark scripts & models
• maas/actions/ ← atomic “agent” behaviors (plugins)
• maas/provider/ ← LLM API adapters & registry
• maas/rag/ ← retrieval-augmented generation engine
• maas/strategy/ ← search-space, planner, solver logic
• maas/tools/ ← tool invocation (email, TTS, browser, ML preprocessing)
• Key config files: config2.yaml, browser_config.py, llm_config.py, redis_config.py, etc.
• Entry point: setup.py + requirements.txt for packaging; examples invoke python -m examples.maas.optimize.
README for additional insights
• Two core components:
a. Agentic Supernet (probabilistic distribution over agent pipelines)
b. Controller that samples sub-networks per query and updates both distribution and operators via feedback.
• External dependencies: downloaded datasets in maas/ext/maas/data; API keys for LLM providers.
• Benchmarks orchestrated by examples/maas/optimize.