maze-agent/Maze

A distributed framework for LLM agents

What it solves

Maze is a distributed framework designed to manage and scale LLM agent workflows. It addresses the challenges of resource contention, task parallelism, and reliability in large-scale agentic systems by providing a structured way to to distribute tasks across a cluster of workers.

How it works

Maze uses a head-worker architecture based on on Ray. It allows developers to define workflows as a series of tasks (static or dynamic) or as ReAct-style agent loops. The framework manages task scheduling using algorithms like FCFS or HACS, and handles resource allocation (CPU, GPU, and I/O) to prevent bottlenecks. It includes a content-addressed artifact store for managing files across distributed nodes without requiring a shared filesystem.

Who it’s for

It is intended for developers building complex, multi-agent systems or high-performance agent workflows that require distributed execution, fine-grained resource management, and high reliability.

Highlights

  • Distributed Execution: Supports standalone or cluster-based deployment with automatic worker recovery and fault tolerance.
  • Task-Level Management: Enables fine-grained control over task parallelism, retries, and timeouts.
  • Heterogeneous Scheduling: Implements specialized queues for GPU, CPU, and I/O tasks to optimize throughput.
  • ReAct Workflow Support: Built-in templates for ReAct agents with support for tool calls, MCP servers, and agent traces.
  • Observability: Provides a unified console (Maze Playground) for drag-and-drop workflow building, run history, and cluster diagnostics.
  • Backend Compatibility: Can serve as a runtime backend for other frameworks, such as LangGraph.