The right open source AI agent framework depends on the job. This guide sorts LangGraph, CrewAI, AutoGen, OpenHands, and open source Kortix by what each one runs, where memory and state live, and who reviews a change before it lands.
What each framework runs, where state lives, and how work is reviewed.
| Criteria | What it is | Open source | What you own | Runs the agents or only defines them | Where memory and state live | Model choice | Human review of changes | The job it fits |
|---|---|---|---|---|---|---|---|---|
| Kortix | The open-source AI Operating System a company runs and owns. | Yes | The whole system: agents, skills, memory, and connectors in your repo. | Runs them: each session on its own cloud computer. | In the repo as files, versioned in the same git history. | Any model with your own keys, or a ChatGPT subscription. | Work reaches main through a change request a person reviews. | Running company work end to end, owned and reviewed. |
| LangGraph | A low-level orchestration framework for long-running, stateful agents. | Yes | Your application code; the framework's runtime stays in it. | Defines and runs orchestration inside your own application. | In the framework's persistence layer, wired into your app. | Any model; docs cover OpenAI, Anthropic, Google. | Human-in-the-loop: inspect and edit state at a chosen point. | Custom stateful runtimes embedded in your own application. |
| CrewAI | A Python framework for role-based agent teams and event-driven flows. | Yes | Your Python app and the crew and flow definitions. | Defines and runs crews inside your Python application. | In CrewAI's memory store, with scopes and vector recall. | Any model you wire in; tools reach any API. | An optional human_input field pauses a task for a person. | Role-based agent teams inside a Python application. |
| AutoGen | A programming framework for multi-agent systems your app hosts. | Yes | Your app and the agent team you compose. | Defines and runs teams your application hosts. | In your app, saved to disk and loaded from storage. | Model clients you add, including OpenAI and others. | A UserProxyAgent pauses the team for approval or feedback. | Research and multi-agent systems your application hosts. |
| OpenHands | A self-hosted front end that runs coding agents. | Yes | The self-hosted stack you run on your own machine. | Runs coding agents on a machine or server you pick. | In each conversation's workspace and persisted state on the host. | Bring your own model: use with any LLM. | You review the conversation and the changes in your repo. | Coding tasks: an agent that develops in your repositories. |
acme-co/ your whole company, as filesagents/ who does the workskills/ how your company does a jobmemory/ what it has learned so farkortix.yaml rules, triggers, connectors, machineAgents, skills, memory, and configuration are files you can read and diff.
Start with the shape of the work, then pick. A single agent loop inside a service you control points to a low-level orchestration framework such as LangGraph. A team of named roles that runs in Python points to CrewAI. Research into multi-agent behavior that you host yourself points to AutoGen. Coding work in your repositories points to OpenHands. Company work that must land as reviewed changes your team owns points to Kortix, which runs each session on its own cloud computer and returns finished work through a change request. The comparison table on this page sets the five side by side.
Most comparisons rank frameworks on features and skip the layer each one sits in. A framework defines agents: prompts, tools, and control flow live in your application, and your team owns the runtime around them. A platform runs agents and holds the company state they work from. Two questions decide the layer. Where does the finished work land, and can you read the history of how it got there? If the answer is inside your application and your own logs, a framework is the right layer. If the answer is a repo your team owns, with every change reviewed before it merges, you are choosing a platform. The open source agent framework alternatives guide sets the frameworks side by side, and framework or platform explains the split.
LangGraph, CrewAI, and AutoGen define and run agents inside an application you build and operate. That is the right layer when the agent is part of your product. When the agent is part of the company, the layer changes. Kortix is the open-source AI Operating System: agents, skills, company memory, and connectors live in one git repo you own, each session runs on its own cloud computer, and session work reaches main through a change request a person reviews. Merge is default-deny for agents. To see the two layers on one stack, read Kortix versus LangGraph.
A framework defines agents inside an application your team builds and runs. A platform runs the agents and holds the state they work from. Kortix is the open-source platform: it runs each session on its own cloud computer and lands work as reviewed changes.
No. Kortix runs agents, skills, memory, and connectors from one git repo you own. You would add a framework only when a specific agent has to live inside your own application.
Match the framework to the job. LangGraph for a custom stateful loop, CrewAI for role-based teams in Python, AutoGen for multi-agent research, OpenHands for coding work. Choose Kortix when the work has to be owned and reviewed.
Yes. Kortix is open source. Read it, fork it, audit it, and run it on Kortix Cloud, in your VPC, or on your own on-prem network.
One repo for the agents, the skills, the memory, and the connectors.