Hey HN,<p>I built Manas to solve challenges I faced in my PhD research on formally verified systems for LLM applications. After repeatedly implementing similar patterns for agent orchestration, I decided to create a reusable framework.<p>What makes Manas different is its approach to multi-agent workflows. It implements a think-act-observe cycle for agents with proper state persistence, while providing visualizable workflow orchestration between specialized agents. The core flow execution logic is formally verified, which was critical for my research.<p>Interestingly, I "vibe coded" much of Manas - using LLMs to generate code while I guided the architecture. This meta approach of using LLMs to build LLM orchestration tools proved surprisingly effective.<p>The framework is provider-agnostic (OpenAI, Anthropic, HuggingFace, Ollama) and includes vector store integrations (FAISS, Chroma, Pinecone) for RAG applications. Everything is async-first with proper error handling.<p>The docs include several examples to get you started: <a href="https://manas.koley.in" rel="nofollow">https://manas.koley.in</a><p>GitHub: <a href="https://github.com/arkokoley/manas">https://github.com/arkokoley/manas</a><p>I would love to hear your thoughts!