Skip to content

About

Awesome list of multi-agent workflow tools, frameworks, platforms, protocols and research

Resources

Readme
License

Contributing

Contributing
Activity

Stars

0 stars

Watchers

0 watching

Forks

0 forks
Report repository
BranchesTags
 
 

Latest commit

 

History

6 Commits

Folders and files

Repository files navigation

Awesome Multi-Agent Workflows Awesome

A curated directory of multi-agent workflows — the frameworks, platforms, protocols, and research for getting multiple AI agents to work together — as of September 2026.

Single-agent frameworks are out of scope: every entry here is about collaboration — crews, swarms, supervisor teams, debates, handoffs, and the protocols and durable infrastructure underneath them.

Scope: This list is about multi-agent teams as working systems. Single-agent frameworks are out of scope; every entry concerns collaboration — crews, swarms, supervisor teams, debates, handoffs — and the products and infrastructure that run them: commercial/SaaS platforms, cloud multi-agent services, and durable execution infrastructure (retries, checkpoints, human-in-the-loop). For the underlying coordination mechanisms (protocols, memory systems, orchestration frameworks in depth, research papers), see awesome-AI-agent-orchestration; for the wider single-agent ecosystem, see awesome-ai-agents.

Verification confidence: every entry is stamped ✅ verified (facts checked on the official docs, repo, or paper page) or ⚠️ unverified (third-party or ambiguous). Capabilities and licenses are never guessed — if the official source is ambiguous, the entry says so. Machine-readable records live in data/multi-agent-workflows.json with a verified boolean per entry. 88 of 89 entries verified on official sources (2026-09-29).

2026 Highlights

  • Agent-to-agent protocols grew up: A2A hit v1.0.0 as a Linux Foundation project (March 2026), IBM's ACP merged into it (Aug 2025), and the MCP spec's 2026-07-28 stateless redesign moved it under the Agentic AI Foundation. The split is clean: A2A is agent-to-agent, MCP is agent-to-tool.
  • AutoGen is in maintenance mode — Microsoft points new framework users to the Microsoft Agent Framework, while Magentic-One remains the research showcase for orchestrator-led teams.
  • Big-cloud multi-agent is GA: Bedrock's supervisor/collaborator pattern, the Gemini Enterprise Agent Platform (Vertex AI Agent Engine rebranded), Microsoft Foundry (Azure AI Foundry renamed) multi-agent workflows, and Agentforce orchestration are all production offerings.
  • Durable execution is the substrate: Temporal's official OpenAI Agents SDK integration, Trigger.dev v3 (Apache-2.0, relaunched Sept 2025 with an AI-agent focus), and DBOS are all courting agent workloads — because agent teams need retries, checkpoints, and human-in-the-loop that survive restarts.
  • License watch: Inngest is SSPL, Windmill is AGPL-3.0, n8n is fair-code (Sustainable Use License), Restate's runtime is BSL-1.1 — none of these are OSI open source, whatever third-party lists claim. LangGraph Platform was renamed LangSmith Deployment (Oct 2025).

Contents

Open-source orchestration frameworks

Libraries and frameworks you embed or self-host, where agents collaborate through crews, graphs, conversations, or role-play.

  • AgentScope — ✅ verified 2026-09-29. Multi-agent framework emphasizing agent teams and leader/worker coordination, planning, and human-agent interaction, with a visual multi-agent platform. Gotchas: Official repository README confirms Apache-2.0 license and describes agent teams, coordination, and planning as core capabilities.
  • Agno — ✅ verified 2026-09-29. Lightweight open-source framework and runtime for building agents, teams, and workflows for production agentic AI, with a runtime environment for serving them. Gotchas: Official repository (agno-agi org) opened; metadata and README confirm Apache-2.0; Formerly Phidata; Actively developed.
  • AutoGen — ✅ verified 2026-09-29. Microsoft Research framework for creating multi-agent AI applications that can act autonomously or work alongside humans; layered design with a Core API (message passing, event-driven agents, distributed runtime) and AgentChat (group chats, two-agent patterns), plus no-code AutoGen Studio. Gotchas: Official repository opened; legal notices state code is MIT (LICENSE-CODE), docs/content are CC-BY-4.0; In maintenance mode as of 2026; Microsoft directs new users to the Microsoft Agent Framework; Magentic-One (a multi-agent team for web/code/file tasks) is implemented in this repo rather than standalone.
  • CAMEL — ✅ verified 2026-09-29. Multi-agent framework supporting communicative agents, agent societies, role playing, task automation, code generation, communication, memory, and data generation/simulation. Gotchas: Official repository (camel-ai org) opened; repository metadata and README confirm Apache-2.0 license; Actively developed (~13k+ stars).
  • CrewAI — ✅ verified 2026-09-29. Framework for building collaborative AI agent crews with defined roles, goals, and memory; Crews handle role-based collaboration while event-driven Flows enable more controlled, deterministic workflows. Gotchas: Official repository confirms MIT license and documents both Crews and Flows; Very active project (~35k+ stars); Official docs linked from the README: https://docs.crewai.com/.
  • Google Agent Development Kit (ADK) — ✅ verified 2026-09-29. Google's open-source agent development kit; code-first framework for building single agents and multi-agent systems with composition primitives such as sequential, parallel, and loop workflow agents plus sub-agent delegation. Gotchas: Python SDK; TypeScript and Go SDKs exist separately.
  • LangGraph — ✅ verified 2026-09-29. Stateful orchestration layer for LangChain agents; models agents as nodes in a graph with durable execution, persistence, and streaming, supporting supervisor, sub-agent, and hierarchical multi-agent patterns. Gotchas: Official LangChain documentation; official README carries an MIT badge; The core multi-agent orchestration primitive in the LangChain ecosystem.
  • MetaGPT — ✅ verified 2026-09-29. Multi-agent framework implementing an SOP-driven 'software company' — role-based agents (product manager, architect, engineer, QA) collaborate to generate PRDs, designs, and code. Gotchas: Official repository (FoundationAgents org) confirmed MIT in repository search results; One of the pioneering role-based multi-agent frameworks.
  • OpenAI Agents SDK — ✅ verified 2026-09-29. Lightweight open-source Python library for building agentic apps; multi-agent workflows are expressed through handoffs and agents-as-tools, with built-in tracing and guardrails. Gotchas: Official OpenAI repository opened; README describes it as a lightweight yet powerful framework for multi-agent workflows, MIT licensed; Supersedes the experimental OpenAI Swarm.
  • PraisonAI — ✅ verified 2026-09-29. Production-oriented multi-agent framework with self-reflection and low-code orchestration of role-based agent crews. Gotchas: Official documentation site opened; MIT badge confirmed in the docs source; Main repo is mervinpraison/praisonaidocs (docs) with the framework under the PraisonAI org.
  • Semantic Kernel — ✅ verified 2026-09-29. Microsoft SDK for building AI agents and orchestrating agentic and multi-agent AI systems; lightweight open-source development kit that lets developers add agents and multi-agent orchestration to applications. Gotchas: Official Microsoft repository; README confirms MIT license and explicitly describes building and orchestrating multi-agent systems; Multi-language (Python, .NET, Java).
  • SmythOS Runtime Environment (SRE) — ✅ verified 2026-09-29. Open-source cloud-native runtime for agentic AI: runtime environment, SDK, and CLI for building, running, and managing agents locally, in cloud, or at the edge, with built-in security and 40+ production-ready components. Gotchas: Official repository (SmythOS org) opened; GitHub metadata confirms MIT license; Repository topics include multi-agent, multi-agent-systems, agent-framework, and orchestration; Newer project (created 2025-06-07, ~1.3k stars).

Commercial / SaaS multi-agent platforms

Hosted products that build, run, and manage multi-agent workflows for you.

  • Bardeen — ✅ verified 2026-09-29. AI agents for go-to-market work: agentic web scraper, AI web search, contact enrichment, and AI lead qualification assembled into reusable playbooks across 100+ apps. Gotchas: Repositioned from personal browser automation to B2B GTM (sales teams); Free tier: 100 credits/mo; Basic $10/mo; Premium $50/mo; SOC 2, GDPR, CASA Tier 2/3.
  • Beam AI — ✅ verified 2026-09-29. Agentic Process Automation platform: pre-built self-learning AI agent suites for finance, HR, customer service, and other ops; AgentOS supports multi-agent collaboration, cloud/on-prem/hybrid deployment. Gotchas: Enterprise-focused, white-glove onboarding ('live in weeks'); claims 10M+ agent tasks processed for Fortune 500 customers; Also listed on AWS Marketplace.
  • CrewAI AMP — ✅ verified 2026-09-29. CrewAI's commercial platform (AMP Cloud hosted, AMP Factory on customer infrastructure) for running multi-agent crews and flows in production. Gotchas: Enterprise product is branded AMP (AMP Cloud hosted, AMP Factory on customer infra), not 'CrewAI Enterprise'; MIT open-source framework; commercial platform with a free tier.
  • Dust — ✅ verified 2026-09-29. 'Multiplayer AI' platform where teams build and deploy AI agents in shared workspaces connected to company knowledge and 100+ tools; no-code agent builder, model-agnostic, enterprise governance. Gotchas: Paris/SF; $40M Series B May 2026 (Abstract, Sequoia), 3,000+ orgs, 300k+ agents deployed; SOC 2 Type II, GDPR, HIPAA-ready.
  • Flowise — ✅ verified 2026-09-29. Visual drag-and-drop builder for agentic systems — multi-agent systems, chat assistants, RAG — deployable as API/SDK/embed; managed Flowise Cloud with Free/Starter/Pro tiers. Gotchas: IMPORTANT: FlowiseAI archived the OSS GitHub repo on 2026-08-13 and wound the project down (support ended 2026-08-31); Flowise Cloud still sells (Free $0, Starter $35/mo, Pro $65/mo as of Sept 2026), but upstream is unmaintained.
  • Gumloop — ✅ verified 2026-09-29. No-code platform where business teams build and orchestrate AI agents and automations across Slack, Teams, and email; model-agnostic with BYOK and enterprise controls (SSO, RBAC, audit logs) via Gumstack. Gotchas: Original visual workflows now labeled 'Legacy' — the product is agent-centric; Pro starts at $37/mo (20k credits); no free plan per Sept 2026 review.
  • Lindy — ✅ verified 2026-09-29. No-code AI teammate that lives in Slack/iMessage and joins meetings; 40+ skills, reusable skills learning, 1,000+ integrations plus MCP, browser control on Pro+. Gotchas: Relaunched Feb 2026 from agent-builder to AI-assistant product; visual builder still exists for enterprise teams; Plans: Plus $29.99/mo, Pro $99.99/mo, Max $199.99/mo, Enterprise custom; $50 free credits, no card.
  • Relevance AI — ✅ verified 2026-09-29. Low/no-code 'AI workforce' platform for building, deploying, and managing teams of specialist agents and multi-agent workflows; Invent 2.0 (launched Sept 29, 2026) auto-builds agent workforces from process interviews. Gotchas: Free tier exists historically; current positioning is enterprise-led (custom pricing); SOC 2 Type II, GDPR; Used by Canva, Autodesk, Confluent.
  • SmythOS — ✅ verified 2026-09-29. Agent operating system: visual Agent Studio builder plus a cloud-to-edge runtime for building and deploying autonomous multi-agent systems; supports MCP client/server and agent-to-agent (A2A) collaboration. Gotchas: Commercial platform with MIT open-source runtime components (SmythOS SRE); Deploy via APIs, webhooks, scheduled agents, chat widgets.

Cloud multi-agent services

First-party multi-agent offerings from the cloud providers — managed orchestration, supervisor patterns, and agent runtimes.

  • Amazon Bedrock Agents — Multi-Agent Collaboration — ✅ verified 2026-09-29. Managed supervisor pattern: designate one Bedrock Agent as supervisor and associate collaborator sub-agents that plan, split tasks, work in parallel, and hand off via SUPERVISOR or SUPERVISOR_ROUTER modes. Gotchas: Part of Amazon Bedrock (AWS); Region-limited (Bedrock Agents availability varies by AWS region); Supervisor must be saved before collaborators can be attached.
  • Cloudflare Agents — ✅ verified 2026-09-29. Stateful agents on Cloudflare Workers (Durable Objects) that can call and coordinate other agents via getAgentByName; official examples include multi-agent workers/pipeline patterns. Gotchas: Agents SDK (MIT) is open source but the runtime is Cloudflare's managed Workers platform; orchestration is code-driven (no visual orchestrator); Multi-agent patterns documented in the official cloudflare/agents examples repo.
  • Databricks Mosaic AI — Supervisor Agent — ✅ verified 2026-09-29. Managed agent platform on Databricks: build multi-agent systems with a Supervisor Agent that orchestrates Genie Agents, agent endpoints, Unity Catalog functions, MCP servers, and custom agents. Gotchas: Docs brand it as 'agents on Databricks' (Mosaic AI Agent Framework); ResponsesAgent (MLflow 3) is the current authoring interface; Supervisor Agent and some features are in beta.
  • Gemini Enterprise Agent Platform (Agent Runtime) — ✅ verified 2026-09-29. Google's managed agent runtime (Vertex AI Agent Engine rebranded) for deploying ADK multi-agent orchestrations — sub_agents, AgentTool, and graph workflows — as a service. Gotchas: ADK itself is Apache-2.0 open source, but the managed runtime is a cloud service.
  • IBM watsonx Orchestrate — ✅ verified 2026-09-29. Enterprise multi-agent collaboration platform: supervisor/orchestrator routes work across skills (independent agents) via Intent Parser and Flow Orchestrator with shared context/memory store. Gotchas: Verified on official IBM page stating watsonx Orchestrate 'makes it easy to enable multi-agent collaboration'; Dedicated docs at ibm.com/docs/en/watsonx-orchestrate (403 for the VM egress, not a product issue); SaaS on IBM Cloud and AWS Marketplace; Agent Connect integrates external-framework agents.
  • Microsoft Copilot Studio — Multi-Agent (Connected Agents) — ✅ verified 2026-09-29. Low-code agent platform where an orchestrator agent uses generative orchestration to delegate to connected specialist agents (Copilot Studio, Foundry, Fabric Data, A2A external agents). Gotchas: Official Learn training module (based on the classic Copilot Studio experience; a new experience shipped in 2026); Multi-agent capability is preview/maturing; overlaps with Foundry workflows.
  • Microsoft Foundry — Multi-Agent Workflows — ✅ verified 2026-09-29. Managed multi-agent orchestration in Foundry: visual/YAML workflows (Sequential, Group chat, Human-in-the-loop) invoking multiple Foundry agents with variables, versioning, and tracing. Gotchas: Renamed from 'Azure AI Foundry' to 'Microsoft Foundry' (2026).
  • Oracle AI Data Platform Workbench — ✅ verified 2026-09-29. Visual canvas builder for multi-agent systems on OCI: a Supervisor Agent node orchestrates executor specialist agents (SQL, RAG, HTTP, MCP tools) with memory and state-isolation policies. Gotchas: Official docs.oracle.com 'Agent Creation' guide covers Multi-agent Systems and Supervisor Patterns; Backed by OCI Generative AI service models; related to (but a distinct, newer builder from) the OCI Generative AI Agents service.
  • Salesforce Agentforce — Multi-Agent Orchestration — ✅ verified 2026-09-29. Primary agent routes tasks to a team of specialist secondary agents powered by the Atlas Reasoning Engine, with A2A support for third-party agents and built-in observability. Gotchas: Verified on the official Salesforce product page (marketing page rather than help docs); GA as of Sept 2026.
  • ServiceNow AI Agent Orchestrator — ✅ verified 2026-09-29. Orchestrator coordinates teams of ServiceNow AI agents via Chain-of-Thought + ReAct, with shared memory across agent teams, AI Agent Studio, AI Agent Fabric (MCP/A2A), and AI Control Tower governance. Gotchas: Verified on official ServiceNow-hosted 'ServiceNow AI Agent Well-Architected Review' whitepaper (April 2026); Now Assist / AI Agents are licensed per ServiceNow instance; GA since the Zurich release.

Agent-to-agent protocols & standards

Protocols and specifications for agents to discover each other, delegate work, exchange messages, and transact.

  • A2A (Agent2Agent Protocol) — ✅ verified 2026-09-29. Open protocol for peer-to-peer agent collaboration: agents publish an Agent Card at /.well-known/agent-card.json, then exchange tasks and messages over JSON-RPC 2.0, gRPC, or REST with SSE streaming and push notifications. Gotchas: Created by Google (Apr 2025), donated to the Linux Foundation / Agentic AI Foundation; v1.0.0 released 2026-03; IBM's ACP was merged into A2A (Aug 2025); Complementary to MCP (agent-to-agent vs agent-to-tool).
  • ACP (Agent Communication Protocol) — ✅ verified 2026-09-29. REST-native OpenAPI-defined protocol for agent-to-agent invocation: Agent Manifest discovery (GET /agents), runs (POST /runs) with sync/SSE/async modes, MIME-typed multipart messages, and trajectory metadata. Gotchas: Launched by IBM Research Mar 2025; merged into A2A under the Linux Foundation (Aug 2025); Reference repo i-am-bee/acp is archived — legacy/superseded, build on A2A instead.
  • Agent Protocol — ✅ verified 2026-09-29. Tech-agnostic OpenAPI specification defining a common REST interface for interacting with AI agents: create tasks (POST /ap/v1/agent/tasks), execute steps, and download/upload artifacts. Gotchas: See the official source for current details..
  • AGNTCY / OASF — ✅ verified 2026-09-29. Cisco Outshift-initiated open collective (now Linux Foundation) building the 'Internet of Agents': OASF is a standardized schema for describing agent capabilities/identity/IO (modeled on OCSF), plus an Agent Directory, SLIM secure messaging, and identity specs. Gotchas: A stack of specs/SDKs rather than one wire protocol; positioned to complement A2A and MCP; Includes Agent Connect Protocol (ACP) for messaging and SLIM (MLS-based, quantum-safe) transports.
  • Agora — ✅ verified 2026-09-29. Decentralized protocol for LLM-agent communication: agents default to natural language for rare exchanges, then negotiate and exchange self-generated protocol documents (hash-identified) to use efficient routines for frequent interactions. Gotchas: Originated from Oxford research; fully decentralized with no central registry; MIT-licensed per the official FAQ.
  • AITP (Agent Interaction & Transaction Protocol) — ✅ verified 2026-09-29. NEAR AI's open protocol for cross-trust-boundary interaction: agents communicate over Chat Threads (largely compatible with the OpenAI Assistants/Threads format) with extensible Capabilities for structured exchanges like payments, forms, and generative UI. Gotchas: Spec in progress; official docs show v0.1.0 Draft; Reference implementation target is agents on the NEAR AI Hub.
  • ANP (Agent Network Protocol) — ✅ verified 2026-09-29. Open protocol suite for the Agentic Web: did:wba decentralized identity, WNS human-readable handles, agent description and discovery documents, secure end-to-end messaging profiles, and application-layer protocols. Gotchas: Spec organized around the ANP 1.1 release line; the meta-protocol specification remains a draft; Governance is a small GitHub community; adoption is thin outside China.
  • AP2 (Agent Payments Protocol) — ✅ verified 2026-09-29. Google's open protocol for agent payment authorization: agents carry cryptographically signed mandates (Intent/Checkout/Payment Mandates as SD-JWT verifiable credentials) so every party can prove a purchase was authorized and audit it later. Gotchas: An authorization/audit layer, not a settlement rail; v0.2.0 (Apr 2026), standardization continuing in FIDO Alliance working groups; Originally an A2A AgentCard extension in v0.1; agent-to-agent delegation is out of scope in the current spec.
  • Coral Protocol — ✅ verified 2026-09-29. MCP-native multi-agent coordination platform: agents join shared sessions over MCP, collaborate via @mention-addressed threads, and can be listed in a registry with Solana-based micropayments to agent developers. Gotchas: More platform + coordination pattern than neutral standard; graph-based multi-agent architecture (Anemoi) rather than hierarchical orchestration.
  • DIDComm — ✅ verified 2026-09-29. Hyperledger/Aries messaging layer for secure agent-to-agent communication: DID-based, end-to-end encrypted, asynchronous messages (transport-agnostic) with standard protocol sequences (DID exchange, credential issuance, present-proof) on top. Gotchas: Genuinely agent-oriented (Aries agents' messaging layer) but rooted in the self-sovereign-identity world; broader multi-agent ecosystem adoption is limited.
  • Eclipse LMOS Protocol — ✅ verified 2026-09-29. Eclipse Foundation protocol built on W3C Web of Things architecture: JSON-LD agent/tool descriptions signed with W3C DIDs, metadata propagation and discovery, transport-agnostic communication (lmosprotocol subprotocol), and DID/OAuth2 security schemes. Gotchas: Protocol text still work-in-progress (v1 namespace published); team signals intent to pursue W3C standardization; Eclipse project (EPL-typical licensing for Eclipse code, unverified).
  • FIPA-ACL / KQML — ⚠️ unverified. The original agent communication languages: KQML (DARPA, 1993) and FIPA-ACL (ratified 2000) standardized speech-act performatives (inform, request, propose, cfp) and interaction protocols like contract-net and subscribe/notify. Gotchas: Historic/legacy: formal semantics and ontology overhead limited adoption; FIPA became an IEEE Computer Society committee in 2005.
  • MCP (Model Context Protocol) — ✅ verified 2026-09-29. Open standard (Anthropic, now under the Agentic AI Foundation) that standardizes how LLM applications connect to external tools, data sources, and prompts via JSON-RPC 2.0 in a client-host-server model. Gotchas: Current spec revision 2026-07-28 (stateless redesign).
  • NANDA — ✅ verified 2026-09-29. MIT Media Lab research infrastructure for the 'Internet of AI Agents': the NANDA Index (federated agent registry/name resolution), AgentFacts (capability documents signed as W3C Verifiable Credentials), and KYA third-party attestation. Gotchas: Discovery and trust layer, not a messaging wire protocol — it sits above MCP and A2A and does protocol translation via the NANDA Adapter.

Workflow & orchestration infrastructure

Durable-execution and workflow engines with a real multi-agent story — the substrate long-running agent teams run on.

  • Activepieces — ✅ verified 2026-09-29. Open-source (Community Edition MIT) visual AI-automation platform: native AI pieces incl. a 'Run Agent' action that reasons through multi-step tasks with tools, human-in-the-loop approval pieces, and 280+ pieces published as MCP servers usable directly by LLMs in Claude Desktop/Cursor/Windsurf. Gotchas: Community Edition is MIT; packages/ee/ (enterprise features) are under a separate Commercial License; Like n8n, more automation-platform than durable agent infra — included on the strength of the Run Agent action + MCP toolkit.
  • Camunda — ✅ verified 2026-09-29. Enterprise process-orchestration platform (Zeebe engine) that added native agentic capabilities: GA'd AI Agent connector (8.8, Oct 2025) with ad-hoc sub-processes where the LLM dynamically picks BPMN activities, A2A client connectors, built-in MCP server, and ProcessOS agent intelligence layer (beta). Gotchas: Enterprise BPMN platform — heaviest fit in this list; engine is source-available under the Camunda License v1 (production requires Enterprise license), not OSS.
  • DBOS — ✅ verified 2026-09-29. Durable-execution library (Python/TypeScript/Go/Java) backed by Postgres: wrap any agent in DBOS workflows to checkpoint every LLM call and tool execution; first-party integrations for Vercel AI SDK (durableCalls middleware), OpenAI Agents SDK (DBOSRunner), Google ADK, and Pydantic AI. Gotchas: MIT across dbos-transact SDK repos; Lightweight approach: no separate orchestration cluster — durability lives in your app + Postgres; Official docs have a dedicated AI section (durable agents, parallelizing/scaling agents, HITL).
  • Hatchet — ✅ verified 2026-09-29. Distributed, fault-tolerant task queue and DAG workflow engine backed by Postgres: named 'AI agents' as an official use case — define agentic workflows as code, retry failures, parallelize agent actions, with fair queuing, rate limits, durable sleeps/event waits, and OpenTelemetry. Gotchas: Fully MIT licensed end-to-end; self-hosted or Hatchet Cloud; Younger project (0.x as of Sept 2026, ~7.9k stars) with fewer production war stories than Temporal; load-tested to ~10k tasks/s.
  • Inngest — ✅ verified 2026-09-29. Event-driven durable-execution platform for TypeScript/Python/Go: write step functions (step.run/step.invoke/step.sendEvent) that look like normal code while Inngest handles retries, state, sleeps and event waits; first-class agent patterns (sync/async/scheduled sub-agent delegation) and durable AI-agent primitives. Gotchas: Self-hostable since Jan 2026, but licensed under SSPL (not OSI open-source); Agent angle documented in Inngest's own blog ('durable AI agents', agentic patterns); Best DX of the durable-execution class; lighter-weight than Temporal.
  • LangSmith Deployment — ✅ verified 2026-09-29. LangChain's managed deployment for LangGraph multi-agent applications (renamed from LangGraph Platform, Oct 2025); the framework itself stays MIT open source. Gotchas: Renamed from LangGraph Platform to LangSmith Deployment in Oct 2025; Not full durable execution like Temporal — checkpointing + replay only.
  • n8n — ✅ verified 2026-09-29. Fair-code visual workflow automation platform with mature AI capabilities: LangChain-based AI Agent nodes (Chat Trigger → AI Agent → chat model + memory + tools), human approvals, 1500+ integrations; self-hostable or n8n Cloud. Gotchas: Source-available under the Sustainable Use License (fair-code, not OSI open-source); More 'AI workflow automation' than durable multi-agent infra — no event-replay durability; HITL is webhook/Wait-node based; Included because its AI Agent nodes are real and widely used for agentic automation.
  • Orkes Conductor — ✅ verified 2026-09-29. Open-source (Apache 2.0) event-driven agentic workflow engine, originally Netflix Conductor and maintained by Orkes: durable execution with native LLM task types (14+ providers), MCP tool calling, vector DB support for RAG, and human approval gates; scales to billions of executions. Gotchas: OSS engine (github.com/conductor-oss/conductor) + Orkes SaaS enterprise offering; Agent features are native task types (not bolt-on); Polyglot workers (Java/Python/Go/JS/C#/Ruby/Rust).
  • Prefect — ✅ verified 2026-09-29. Python-native workflow orchestration (@flow/@task decorators, retries, caching, scheduling): widely used as the orchestration layer under AI agents and agentic coding pipelines — durable per-agent runs with task-level observability, often paired with frameworks like Pydantic AI. Gotchas: Core engine Apache-2.0 (note: /server directory is under the Prefect Community License).
  • Restack — ✅ verified 2026-09-29. Enterprise AI agent platform and backend framework for building and deploying reliable AI agents: durable workflow orchestration, long-running agent execution, auto-generated REST APIs per agent/workflow, MCP tool exposure, Python/TypeScript SDKs on Kubernetes. Gotchas: Proprietary platform (enterprise tier); orchestration layer is built on Temporal under the hood; Primarily commercial — open-source examples and SDKs, but the core platform is not OSS.
  • Restate — ✅ verified 2026-09-29. Open durable-execution foundation (Rust runtime + TS/Java/Python/Go/Kotlin SDKs) for building resilient agents: journals every step for crash recovery, durable promises for human-in-the-loop approvals, rollback/saga patterns for agents, and durableCalls middleware for Vercel AI SDK / LangChain / Pydantic AI. Gotchas: Runtime is BSL 1.1 source-available (not OSI open-source); SDKs are MIT; Dedicated AI docs (durable agents, HITL approvals, rollbacks); Runs as a server in front of your agent service — similar footprint to a message broker.
  • Temporal — ✅ verified 2026-09-29. Open-source durable-execution workflow engine (workflows/activities/signals); a first-class runtime for long-running multi-agent systems — official OpenAI Agents SDK integration (temporalio.contrib.openai_agents), framework-agnostic multi-agent fleets (Google ADK + LangGraph on Temporal), and durable human-in-the-loop signals. Gotchas: Server is MIT (Go); SDKs vary; Agent integration evidence: official temporalio/sdk-python openai_agents plugin and Temporal's Aug 2026 multi-agent blog; Known pitfall from docs: separate task queues for slow LLM inference vs other activities to avoid starvation.
  • Trigger.dev — ✅ verified 2026-09-29. Open-source (Apache 2.0) developer platform for building AI agents and long-running durable workflows in TypeScript: no timeouts, durable checkpointing, automatic retries, queues/concurrency, human-in-the-loop waitpoints, tool calling/MCP support, and an MCP server for AI coding assistants. Gotchas: Originally a background-jobs framework (HN 2023); relaunched v3 in Sep 2025 with explicit AI-agent focus (YC W23, ~15.2k stars); Self-hostable via Docker Compose/Kubernetes or managed cloud.
  • Windmill — ✅ verified 2026-09-29. Open-source developer platform/workflow engine that turns scripts (Python, TS, Go, Bash, SQL...) into workflows, webhooks and UIs: DAG flows with branching, retries, approval steps (human-in-the-loop), and a built-in MCP server that exposes workflows to AI agents. Gotchas: AGPLv3 core (copyleft — check implications for hosted use; some modules Apache-2.0); ~16k stars; Strong fit for script-centric agent tool execution, approval gates, and agent-triggered flows via MCP.

Research papers & systems

Published papers and their accompanying systems on multi-agent collaboration, debate, and emergent behavior.

  • AgentLite — ✅ verified 2026-09-29. Salesforce AI Research's lightweight library for building task-oriented LLM agents and experimenting with reasoning strategies and multi-agent architectures. Gotchas: Paper + repo (SalesforceAIResearch/AgentLite); using arXiv URL.
  • AgentVerse — ✅ verified 2026-09-29. Multi-agent framework that dynamically recruits and adjusts agent team composition during task-solving, and studies emergent positive/negative social behaviors of agent groups (ICLR 2024). Gotchas: Paper + repo (OpenBMB/AgentVerse); using arXiv URL.
  • AutoAgents — ✅ verified 2026-09-29. Automatically generates and coordinates a team of specialized agents adapted to the task at hand, with an observer role reflecting on plans and agent responses to improve coherence. Gotchas: Paper + repo (Link-AGI/AutoAgents); using arXiv URL.
  • AutoGen (paper) — ✅ verified 2026-09-29. Microsoft framework for building LLM applications through multi-agent conversation: 'conversable agents' combine LLMs, tools, and human input orchestrated via conversation programming. Gotchas: Paper + repo (microsoft/autogen); using arXiv URL.
  • CAMEL: Communicative Agents — ✅ verified 2026-09-29. Role-playing framework where two LLM agents (AI User and AI Assistant) autonomously cooperate via 'inception prompting' to complete complex tasks; foundational work on communicative agents (NeurIPS 2023). Gotchas: Paper + open-source repo (camel-ai/camel, Apache 2.0 for code); using arXiv URL.
  • ChatDev — ✅ verified 2026-09-29. Virtual software company: role-playing agents (CEO, CTO, programmer, reviewer, tester) collaborate through a 'chat chain' with communicative dehallucination to build complete software (ACL 2024). Gotchas: Paper + repo (OpenBMB/ChatDev, Apache 2.0 code / CC BY-NC 4.0 data); using arXiv URL.
  • DyLAN — ✅ verified 2026-09-29. Dynamic LLM-agent network: automatically selects an agent team from a candidate pool and deactivates underperforming agents mid-task via an LLM ranker, with early stopping for efficiency. Gotchas: Paper + repo (SALT-NLP/DyLAN); using arXiv URL.
  • Generative Agents — ✅ verified 2026-09-29. Stanford 'Smallville' simulation: 25 memory-enabled LLM agents exhibit believable emergent social behavior (information diffusion, relationship formation) in a sandbox world (UIST 2023). Gotchas: A social simulation rather than a workflow framework; included for its multi-agent social-behavior relevance and influence on later multi-agent simulation work.
  • Magentic-One — ✅ verified 2026-09-29. Microsoft Research's generalist multi-agent system: a lead Orchestrator plans, delegates to specialist agents (web browsing, file handling, coding), and recovers from errors across GAIA, AssistantBench, and WebArena. Gotchas: Research system with open-source implementation in microsoft/autogen; using arXiv URL.
  • MetaAgents — ✅ verified 2026-09-29. Collaborative generative agents with human-like reasoning and specialized skills coordinate in a simulated job-fair environment to study team formation and task-oriented multi-agent coordination. Gotchas: Latest arXiv revision (2025) retitles it 'MetaAgents: Large Language Model Based Agents for Decision-Making on Teaming'; same paper, same ID.
  • MetaGPT (paper) — ✅ verified 2026-09-29. Encodes human Standard Operating Procedures (SOPs) into multi-agent collaboration: specialized roles exchange structured artifacts via a shared message pool to build software (ICLR 2024 oral). Gotchas: Paper + repo (geekan/FoundationAgents MetaGPT, MIT); using arXiv URL.
  • Multiagent Debate — ✅ verified 2026-09-29. Multiple LLM instances independently propose answers, then debate their reasoning over several rounds to converge on a more factual and better-reasoned final answer (ICML 2024). Gotchas: Founding paper of the multi-agent-debate sub-literature; the 'MultiAgent Debate (Du et al.)' entry merged here.

Benchmarks & evals

Benchmarks that measure multi-agent collaboration, coordination, negotiation, and competition.

  • AucArena — ✅ verified 2026-09-29. Evaluation suite simulating auctions where LLM bidding agents compete under budgets, benchmarking strategic planning and execution (budget management, goal adherence, adaptive strategy) in dynamic competitive multi-agent scenarios. Gotchas: Official suite name is 'AucArena' (paper: 'Put Your Money Where Your Mouth Is...'); third-party lists call it 'AuctionArena'.
  • BattleAgentBench — ✅ verified 2026-09-29. Turn-based Battle City wargame benchmark with 7 sub-stages across 3 difficulty tiers that separately assess single-agent navigation, paired-agent task execution, and full multi-agent (2-team, 4-team, 8-agent) cooperation and competition. Gotchas: Tsinghua KEG, arXiv 2024; Level 1 is single-agent — multi-agent relevance is Levels 2–3.
  • Collab-Overcooked — ✅ verified 2026-09-29. Two LLM agents cooperate via natural language in an extended Overcooked-AI kitchen across 30 open-ended tasks at 6 complexity levels, evaluated with process-oriented collaboration metrics (initiate vs. respond, trajectory efficiency). Gotchas: EMNLP 2025 Main; collaboration-focused (not competitive); open-source code at the repo linked in the paper.
  • GameEval — ✅ verified 2026-09-29. Goal-driven conversational games (Ask-Guess, SpyFall, TofuKingdom) where LLM players take distinct roles with cooperative or adversarial objectives, probing cooperative/adversarial strategy, deception, multi-hop reasoning, and long-term planning. Gotchas: Code on GitHub is a third-party mirror (jordddan/gameeval); the paper's official code was released anonymously; license not confirmed.
  • LLM-Coordination (LLM-Co) — ✅ verified 2026-09-29. Pure-coordination multi-agent benchmark: LLM agents act as proactive participants in four coordination games (Agentic Coordination), plus 198 multiple-choice CoordQA questions testing environment comprehension, Theory-of-Mind reasoning, and joint planning. Gotchas: UCSC, arXiv 2023 (v3 2025); code at github.com/eric-ai-lab/llm_coordination per the paper (license not confirmed).
  • MindAgent — ✅ verified 2026-09-29. Gaming-interaction infrastructure for multi-agent planning and coordination, with the CUISINEWORLD benchmark dispatching an LLM coordinator and multiple agents to play simultaneously, scored by the CoS (collaboration efficiency) auto-metric. Gotchas: NAACL 2024 Findings; also covers human-NPC collaboration; code at github.com/mindagent/mindagent (license not confirmed).
  • MultiAgentBench — ✅ verified 2026-09-29. Multi-agent-native benchmark on the MARBLE (Multi-Agent Coordination Backbone with LLM Engine) framework spanning 6 domains — research proposals, Minecraft building, coding, database error analysis (collaborative) plus Werewolf and bargaining (competitive) — with milestone-based KPIs comparing coordination topologies (star/chain/tree/graph). Gotchas: ACL 2025 (arXiv:2503.01935); original repo ulab-uiuc/MARBLE, also mirrored at MultiagentBench/MARBLE; repo license not confirmed.
  • SOTopia — ✅ verified 2026-09-29. Open-ended social-interaction environment where LLM agents role-play characters with private goals across 90 social scenarios (negotiation, collaboration, competition), scored on goal completion and 7 social dimensions via SOTopia-Eval. Gotchas: ICLR 2024 spotlight; dyadic by default, SOTopia-S4 extends to multi-party; requires LLM API key + Redis to run.

Coordination, memory & state

Shared memory, message layers, checkpointing, and state tooling that agents use to coordinate.

  • AgentMail — ✅ verified 2026-09-29. Email infrastructure built specifically for AI agents: programmatic inboxes (thousands of unique addresses via API), full send/receive, native threading, allowlists/blocklists, webhooks + WebSockets for event-driven agent communication. Gotchas: Commercial hosted product (official site agentmail.to; free tier with 3 inboxes); Coordination angle: gives each agent its own identity for async two-way messaging.
  • Cognee — ✅ verified 2026-09-29. Open-source memory control plane for agents: an Extract-Cognify-Load pipeline builds a self-hosted typed knowledge graph (Kuzu/LanceDB/SQLite) plus vector + relational stores so agents share persistent long-term memory across sessions. Gotchas: Apache-2.0; Emphasizes cross-agent knowledge sharing, tenant isolation, traceability/audit; ~30k stars; Commercial Cognee Cloud tier exists.
  • Graphiti — ✅ verified 2026-09-29. Zep's open-source temporal knowledge-graph engine for AI agents: incrementally builds bi-temporal entity/fact graphs (Neo4j, FalkorDB, Neptune) with hybrid semantic + BM25 + graph-traversal retrieval, no LLM calls at query time. Gotchas: See the official source for current details..
  • LangGraph Checkpointing — ✅ verified 2026-09-29. Built-in state persistence for LangGraph agent workflows: a checkpointer snapshots full graph state after every super-step into threads (SQLite, Postgres, Redis, MongoDB), enabling pause/resume, time-travel debugging, and human-in-the-loop. Gotchas: Not a standalone repo — a LangGraph feature (LangGraph itself is MIT); Companion BaseStore provides cross-thread shared memory (PostgreSQL/Redis/in-memory) used by LangMem.
  • LangMem — ✅ verified 2026-09-29. LangChain's official long-term memory SDK for LangGraph agents: extracts and consolidates semantic, episodic, and procedural memories (incl. self-improving system-prompt optimization) backed by LangGraph's BaseStore. Gotchas: MIT; Extraction/consolidation layer over LangGraph storage, not a standalone DB; namespace isolation for multi-tenant memory; agent-facing manage/search memory tools.
  • Letta — ✅ verified 2026-09-29. Platform for stateful AI agents with self-improving memory (formerly MemGPT, from the UC Berkeley research project): OS-inspired hierarchical memory (core blocks + archival recall) so agents persist and compound knowledge across sessions. Gotchas: Apache-2.0 confirmed via GitHub repo metadata.
  • Mem0 — ✅ verified 2026-09-29. Universal memory layer for AI agents: extracts facts from conversations into user/agent/session-scoped long-term memory with hybrid semantic + keyword + entity retrieval; the standard drop-in memory API for multi-agent apps. Gotchas: Apache-2.0 OSS repo plus managed Mem0 Platform (commercial); Self-hostable with 20+ vector-store backends; optional graph backends (Neo4j, Memgraph, Kuzu).
  • Redis Agent Memory — ✅ verified 2026-09-29. Redis's managed memory layer for AI agents (part of Redis Iris, their real-time context engine): two-tier model with session memory (configurable TTL) and background promotion of extracted facts to long-term vector-backed memory. Gotchas: Apache License 2.0 per official README on the repo (opened 2026-09-29); Repo is the research/open foundation; the production distribution is Redis Agent Memory in Iris via Redis Cloud; Ships MCP server interface.
  • SuperMemory — ✅ verified 2026-09-29. Memory and context engine + API for the AI era: embeddings-based hybrid RAG + fact memory + evolving user profiles, with a hosted platform and a self-hostable local binary. Gotchas: MIT (repo LICENSE); Cloud-first with self-host local binary; official plugins for Claude Code, Cursor, Codex, OpenCode, OpenClaw + MCP server at mcp.supermemory.ai/mcp; Note: v0.0.7+ release notes mention a 10,000-document cap on self-hosted ('supermemory lite').
  • Zep — ✅ verified 2026-09-29. Commercial agent-memory platform that stores agent memory as temporal knowledge graphs — facts carry validity intervals so agents track what was true vs. what is true now. Gotchas: Managed product (Zep Cloud) is commercial/closed; open-source Community Edition deprecated April 2025 in favor of Graphiti; Treat acquisition claim as unconfirmed.

Retired / pivoted

Projects that shut down, rebranded, or pivoted away from multi-agent workflows. Newest first; all reported Sept 2026 unless noted.

  • Relay.app — Shut down Sept 14, 2026 (reported).
  • Fixie.ai — Wound down; team moved to Ultravox (reported Sept 2026).
  • MultiOn — Rebranded as AGI, Inc. (reported Sept 2026).
  • Wordware — Pivoted from agent IDE to consumer product 'Sauna' (reported Sept 2026).
  • Cognosys / Ottogrid — Rebranded Ottogrid, then acquired by Cohere (reported Sept 2026).
  • TaskWeaver (Microsoft) — Archived by Microsoft.
  • SuperAGI — Inactive since Jan 2025 (reported).
  • AutoGPT — Original OSS project stalled; the active platform is Polyform Shield (reported).

Guides

Related repositories

Contributing

Entries and corrections are welcome — see CONTRIBUTING.md. Every PR is checked by CI: lychee link check over all markdown files, and validation of data/multi-agent-workflows.json (including the verified boolean and required source_url for verified entries).

License

MIT © 2026 dakotac1994

About

Awesome list of multi-agent workflow tools, frameworks, platforms, protocols and research

Resources

Readme
License

Contributing

Contributing
Activity

Stars

0 stars

Watchers

0 watching

Forks

0 forks
Report repository

Releases

Packages

Contributors