Orchestrator Directory

TAXONOMY

Directory Categories

Every orchestrator falls into one or more of these functional categories, classified by primary use case and architectural approach.

⚙️

General-Purpose Frameworks

4 orchestrators

Flexible foundations for building any multi-agent system, offering primitives for routing, handoffs, and agent lifecycle.

💻

Code Generation & Dev

3 orchestrators

Specialized in collaborative software development: planning, coding, reviewing, and testing through agent teams.

🔬

Research & Analysis

2 orchestrators

Multi-agent pipelines designed for deep research, data synthesis, and analytical workflows across knowledge domains.

🏢

Enterprise & Production

3 orchestrators

Battle-tested frameworks with observability, governance, and compliance features built for production workloads.

🧩

Graph-Based Orchestration

2 orchestrators

Stateful, graph-structured workflows where agents form nodes and edges define handoffs, branching, and cycles.

🎭

Role-Play & Simulation

3 orchestrators

Agents adopting personas, debating, and collaborating in simulated environments to solve complex open-ended problems.

REGISTRY

Twelve frameworks shaping the multi-agent landscape. Each entry is maintained by the community with verified metadata.

CrewAI

crewAIInc

Mature

Role-based multi-agent orchestration in Python. Agents are defined with roles, goals, and backstories, then organized into crews that collaborate on sequential or parallel tasks.

PythonRole-BasedTask PipelineRAG

⭐ 25.8kLicense: MITSince: 2023

AutoGen

Microsoft

Mature

Microsoft's multi-agent conversation framework. Agents engage in flexible dialogue patterns with human-in-the-loop support, code execution sandboxes, and customizable conversation flows.

PythonConversationalCode ExecHITL

⭐ 38.2kLicense: CC-BY-4.0Since: 2023

LangGraph

LangChain

Mature

Graph-based orchestration for stateful, multi-actor applications. Agents are nodes in a directed graph with conditional edges, enabling cycles, branching, and persistent state.

PythonJS/TSGraphStateful

⭐ 10.4kLicense: MITSince: 2024

MetaGPT

DeepWisdom

Growing

Multi-agent meta-programming framework. Assigns SOPs to agents mimicking a software company structure—product manager, architect, engineer, QA—to tackle complex development tasks.

PythonSOP-DrivenSoftware Dev

⭐ 46.1kLicense: MITSince: 2023

ChatDev

OpenBMB

Growing

Virtual software company powered by communicative agents. Simulates the entire software lifecycle through chat-based collaboration between CEO, CTO, programmer, and tester agents.

PythonChat-BasedSoftware Dev

⭐ 26.3kLicense: Apache 2.0Since: 2023

🐫

CAMEL

CAMEL-AI

Growing

Communicative Agents for "Mind" Exploration. A research-focused framework for studying cooperative and competitive behaviors in multi-agent systems through role-playing conversations.

PythonResearchRole-PlaySimulation

⭐ 6.2kLicense: Apache 2.0Since: 2023

Swarm

OpenAI

Experimental

Lightweight, ergonomic multi-agent orchestration. Focused on simplicity with agent handoffs and function-calling. Educational reference implementation, not intended for production use.

PythonLightweightHandoffs

⭐ 18.7kLicense: MITSince: 2024

Agency Swarm

VRSEN

Growing

Agent creation framework built on OpenAI Assistants API. Provides a structured agency paradigm where agents communicate through a managed send-message interface with shared state.

PythonOpenAI APIAgency Model

⭐ 3.8kLicense: MITSince: 2024

Semantic Kernel

Microsoft

Mature

Enterprise AI orchestration SDK supporting C#, Python, and Java. Integrates LLM capabilities as plugins with planners that decompose goals into orchestrated steps across agents.

C#PythonJavaEnterprise

⭐ 22.5kLicense: MITSince: 2023

Letta (MemGPT)

Letta AI

Growing

Stateful agent framework with self-editing memory. Agents maintain persistent memory across conversations and can orchestrate sub-agents, enabling long-running, context-aware workflows.

PythonMemoryStatefulLong-Running

⭐ 13.2kLicense: Apache 2.0Since: 2023

Haystack

deepset

Mature

End-to-end LLM orchestration for building RAG pipelines and agent systems. Component-based architecture with pipelines that connect retrievers, generators, and agent loops.

PythonRAGPipelinesProduction

⭐ 18.3kLicense: Apache 2.0Since: 2020

OpenHands

All Hands AI

New

Autonomous software engineering platform with multi-agent architecture. Agents collaborate on coding, browsing, and command execution within sandboxed environments to resolve issues end-to-end.

PythonSWE AgentSandboxedAutonomous

⭐ 42.6kLicense: MITSince: 2024

MATRIX

Feature Comparison

Side-by-side capabilities across the major orchestrators. Updated quarterly by community maintainers.

Framework	Languages	Graph Workflows	Shared Memory	Human-in-Loop	Streaming	Observability	Sandboxing	Multi-Modal
CrewAI	Python	◐	✓	✓	✓	✓	✗	◐
AutoGen	Python	✓	✓	✓	✓	◐	✓	✓
LangGraph	Python, JS/TS	✓	✓	✓	✓	✓	✗	✓
MetaGPT	Python	◐	✓	✗	✗	✗	◐	✗
ChatDev	Python	✗	✓	✗	✗	✗	✗	✗
CAMEL	Python	✗	◐	◐	✗	✗	✗	✓
Swarm	Python	✗	✓	✗	✓	✗	✗	✗
Semantic Kernel	C#, Python, Java	✓	✓	✓	✓	✓	◐	✓
Letta	Python	◐	✓	✓	✓	◐	✗	◐
Haystack	Python	✓	◐	◐	✓	✓	✗	✓
OpenHands	Python	◐	✓	✓	✓	✓	✓	✓

STANDARDS

Submission Criteria

To be listed in this directory, a multi-agent orchestrator must meet the following baseline requirements.

Open Source

The project must be publicly available under an OSI-approved license. Proprietary or source-available-only projects are excluded from the core directory.

Multi-Agent Architecture

Must support orchestrating two or more autonomous agents that interact, delegate, or collaborate. Single-agent tool-use frameworks do not qualify.

Active Maintenance

At least one commit in the past 90 days and responsive issue triage. Abandoned projects are moved to an archive section after 6 months of inactivity.

Documentation

Minimum viable documentation: a quickstart guide, API reference, and at least one end-to-end example demonstrating multi-agent orchestration.

Community Traction

500+ GitHub stars or equivalent community signal. Newer projects may qualify with a strong technical paper or notable organizational backing.

Reproducible Examples

At least one working example that can be cloned and run with minimal setup. Docker-based or notebook-based demos are preferred for accessibility.

BLIND SPOTS

Gap Analysis

Structural gaps in the current multi-agent orchestrator ecosystem, identified through community surveys and technical audits.

Critical Gap

Cross-Framework Interoperability

No standard protocol for agents built in different frameworks to communicate. An agent in CrewAI cannot hand off to a LangGraph agent without custom glue code. The ecosystem needs an Agent Protocol standard.

Critical Gap

Evaluation & Benchmarking

No shared benchmarks for multi-agent performance. Each framework uses ad-hoc demos. We lack standardized tasks, metrics, and leaderboards to objectively compare orchestration quality.

Important Gap

Production Observability

Most frameworks have minimal tracing. Multi-agent debugging requires understanding which agent said what, when, and why. OpenTelemetry integration is rare and inconsistent.

Important Gap

Cost Control & Token Budgets

Multi-agent systems amplify LLM costs dramatically. Very few orchestrators offer built-in token budgeting, cost caps, or automatic model tier selection per agent role.

Important Gap

Safety & Guardrails

Agent-to-agent communication introduces novel attack surfaces. Prompt injection can cascade across agents. Few frameworks have built-in guardrails or inter-agent trust boundaries.

Moderate Gap

Non-Python Ecosystem

11 of 12 listed frameworks are Python-first. The JavaScript/TypeScript, Go, Rust, and JVM ecosystems have minimal multi-agent tooling, limiting adoption in diverse tech stacks.

WHAT'S MISSING

Community Roadmap

What the ecosystem needs next, prioritized by community vote. This is a living document — the future is shaped by contributors.

Achieved

Framework Cataloging

Initial directory with 12+ frameworks, categorized and documented. Community contribution pipeline established.

Achieved

Feature Comparison Matrix

Side-by-side capability tracking across 24 features. Quarterly update cadence with community PRs for accuracy.

In Progress

Agent Protocol Standard

Defining a universal agent communication protocol so frameworks can interoperate. Draft RFC in review with contributors from 5 frameworks.

Planned

Multi-Agent Benchmark Suite

Standardized evaluation tasks: collaborative coding, research synthesis, debate, and planning. Reproducible harness with public leaderboard.

Planned

Observability Standard

OpenTelemetry-compatible tracing spec for multi-agent systems. Semantic conventions for agent spans, message events, and tool calls.

Future

Safety & Trust Framework

Guidelines and tooling for inter-agent trust boundaries, prompt injection defenses, and cascading failure prevention in production deployments.

Future

Polyglot Agent SDK

Reference implementations of the Agent Protocol in TypeScript, Go, Rust, and Java—breaking the Python monoculture and enabling true cross-language agent networks.

Multi-Agent Orchestrators

Directory Categories