What is pi-multi-agent?

Pi Multi-Agent is a TypeScript framework for building systems where multiple AI agents work together on a single task. Rather than simple prompt chaining, it implements a full lifecycle: a goal is submitted, a planning step breaks it into a dependency graph of sub-tasks, a cluster of specialized agents executes those tasks with real tool calls, an evaluator scores the result on multiple dimensions, and the system replans and retries if the quality threshold is not met. The framework offers six ways agents can collaborate: they can work in sequence, in parallel, through a debate-and-consensus process, as a pool of domain experts, in a creator/critic pair, or in a hierarchy where a supervisor delegates to subordinates. Separately, it offers six communication structures that control how information flows between agents. These two dimensions can be combined depending on what a task requires. Agents have access to a shared memory system that lets them pass data between sessions and build on each other's outputs. Built-in tools include web search, a data analyzer, a web scraper, a code executor, and a report writer. A real-time web dashboard, connected over WebSocket, shows agent status, active tool calls, and progress across the cluster. The framework is written in TypeScript with strict mode throughout and uses DeepSeek as the default language model, configurable to any OpenAI-compatible endpoint. The core classes, DeepPlanner, AgentCluster, and LLMAgentCollaboration, are the main entry points. A quick-start example in the README shows how to produce a 30,000-word research report by decomposing the goal, running up to eight specialized agents, and iterating until a quality score threshold is met. Installation is through npm. Node.js 18 or later and an API key for a supported language model are the only prerequisites.