Agent orchestration decides what happens between the prompt and the finished work: which model runs, which tools are reachable, where state lives, and who reviews the result. LangGraph, CrewAI, AutoGen, and Kortix each answer those four questions differently. Kortix is the open-source AI Operating System that runs and governs the agents.
Rival details come from each project's own repository or documentation, checked October 2026.
| Criteria | Open source and self-host | How work is modelled | Where run state lives | What resumes after a failure | Where company memory lives | Who reviews the output before it lands | Model choice | The job it fits |
|---|---|---|---|---|---|---|---|---|
| Kortix | Open source. Read it, fork it, audit it. Self-host is free. | Triggers start sessions on real machines; each session runs one agent. | In the git repo and the session's own branch. | A sleeping session wakes on demand; its sandbox and branch persist. | Plain files in the git repo you own. | Work reaches main through a change request a person reviews. | Any model provider with your own keys, or your ChatGPT plan. | Agents doing finished work, reviewed by people, from one repo. |
| LangGraph | Open source; a separate deployment platform scales it. | A graph of nodes and edges, run inside your application. | A checkpointer persists each thread's graph state. | Durable execution resumes from exactly where it left off. | A checkpointer for one thread plus a store across threads. | Interrupts pause the graph for human input mid-run. | Any model your code calls, through LangChain integrations. | Long-running, stateful workflows where the graph is the product. |
| CrewAI | Open source; a separate commercial control plane is optional. | Crews of role-based agents plus event-driven Flows. | Flow state, shared between tasks, with a unique id per run. | Saved flow state resumes the flow at its start points. | A unified Memory class, organized in hierarchical scopes. | A human review step you add to a task. | The model you configure per agent. | Role-based multi-agent workflows in Python. |
| AutoGen | Open source; now in maintenance mode. | Agents and Teams on a message-passing runtime. | Agent and team state saved with save_state and loaded back. | You save and load state to continue across restarts. | A memory capability you add to an agent. | Agents act autonomously or work alongside people. | Model clients for OpenAI and Azure; extensions add more. | Research prototypes; new work moves to its successor. |
triggers: - slug: daily-digest type: cron cron: "0 0 9 * * 1-5" timezone: Europe/Berlin prompt: Summarize yesterday's commits and open change requests.A trigger starts a session with no person present, renders its prompt, and sends it to a session.
Four things decide what an agent run becomes, and every approach places them differently. The model decides what reasons and what it costs. The connectors decide what the agent can reach. The run state is what survives a restart, a crash, or a long wait. The review decides who looks at the result before it counts. Most framework comparisons rank features and skip these four, but they are what you inherit when a demo becomes a job.
LangGraph models a run as a graph and persists it through a checkpointer for one thread and a store across threads. CrewAI models a run as crews of roles plus event-driven Flows, with state you choose where to keep and a memory class you configure. AutoGen models a run as agents and teams passing messages, with state you save and load yourself. In all three the prompts, the run environment, and the review wiring stay in your application, and the project's own repository documents the primitives it hands over. The framework or platform guide works through that boundary.
Kortix is the open-source AI Operating System that runs and governs the agents. It keeps the parts a framework hands back in one place: a trigger, a person, or an API call starts a session, and each session runs one agent on its own cloud computer and its own branch. Agents, skills, and company memory are files in one git repo you own, and the model is a choice per agent, per session, or per message, with any model provider and your own keys. Session work reaches main through a change request a person reviews, and merge is default-deny for agents. The how to run an agent framework page covers the path from install to merge.
Multi-agent orchestration is how several agents share one job: which one runs when, what each may reach, where state is kept between steps, and who reviews the result. A framework hands you primitives for those decisions, and Kortix runs the sessions and keeps the state in one repo.
A framework runs wherever you deploy your application. Kortix runs each session on its own cloud computer, so no laptop has to stay open, and self-host is free on your own infrastructure.
Kortix gives each session its own sandbox and its own branch, and the branch name is the session id. Work reaches main only through a change request a person reviews, so two sessions never edit the same branch.
A framework orchestrates the agents inside one application your team builds and operates. Running agents for a whole company, with shared memory, one review gate, and an audit trail, is the layer Kortix adds as the open-source AI Operating System.
One repo for the agents, the skills, the memory, and the connectors.