Distributed AI Platform for Multi-Agent Orchestration

Agentfab is a distributed platform designed for the orchestration of specialized AI agents, where they collaborate through task graphs, persistent knowledge, and bounded review loops. It allows users to define agents in YAML along with their corresponding models, tools, and knowledge bases. By decomposing requests into a Directed Acyclic Graph (DAG), agentfab can efficiently schedule agents to work in parallel, run review loops with strict limits, and maintain state through a robust control plane while also sandboxing shell commands at the operating system level.

The platform facilitates project initiation by breaking down ideas into a task graph, dispatching specialized agents for execution, and writing pertinent artifacts to a data directory. Agentfab fabric combines a Conductor and configuration-defined specialist agents, streamlining the orchestration process. Users can set up a default fabric that includes a conductor, architect, designer, and developer, or customize it as needed. Agents run efficiently within a single process, offering the fastest development loop while supporting deployment across virtual machines, containers, and Kubernetes clusters.

With its streamlined local setup requiring minimal configuration and support for manual production-like deployments utilizing separate processes for the control plane and node hosts, agentfab aims to optimize AI agent interactions. It provides flexibility with additional runtime modes and essential run flags for debugging, external nodes, and data handling, all backed by the capabilities of Go 1.24 and the MIT license.