Open source autonomous AI agents: where the rules live
Four projects come up for open source autonomous AI agents: OpenHands, OpenClaw, Hermes Agent and Kortix. They are four different kinds of tool. What separates them is whether the agent's rules, memory and connectors are files your team can open, and whether a person merges what a run produced. Kortix is the open-source AI Operating System.
Which open-source agent keeps its rules where your team can read them?
Where an agent keeps its rules decides who can run it. A team whose work is coding agents, from one self-hosted console, fits OpenHands, which ships as Agent Canvas. One person who wants an assistant in the chat apps they already use, on their own hardware, fits OpenClaw and its single Gateway. A shell-first solo operator, who wants a terminal interface and a built-in cron schedule, fits Hermes Agent. A company that wants its agents, their skills and its own memory kept as ordinary files in a repository, with a person merging what a run produced, fits Kortix.
Four open-source agent tools, and the place each one keeps the rules a run follows.
kortix.yaml, agents, skills, memoryhermes model switches itEdit the files, push the branch, open the change request
git checkout -b tighten-triage# edit agents/, skills/ or kortix.yamlgit commit -am "Tighten the triage prompt"git push origin tighten-triagekortix cr open --head tighten-triage --title "Tighten the triage prompt"The last command files the branch for review, and the diff is what a reviewer reads.
Four tools, four places to look
A team that has decided to run autonomous agents on its own infrastructure usually arrives with a shortlist from an AI answer: OpenHands, OpenClaw, Hermes Agent, and the names around them. All four are open source, and all four start a run with nobody at the keyboard. The four resemble each other at first glance. One is a console for coding work. One is an assistant in your chat apps. One lives in a terminal. One is aimed at a company that wants the agent's rules under version control. The question that sorts them is where the rules, the memory and the connectors live, and who signs off on what a run produced. Those two facts decide whether the tool stays readable a year after the person who set it up has left.
OpenHands: a control center for coding agents
OpenHands ships as Agent Canvas and describes itself as "the self-hosted developer control center for coding agents and automations" (OpenHands). The product is a front end for agents you may already run. It starts the OpenHands agent out of the box, and it can drive Claude Code, Codex, Gemini, Pi, OpenCode or any ACP-compatible agent across local, remote and cloud backends.
The work it organizes is coding work. An automation publishes a report to Slack, or splits a GitHub issue into tasks, on a schedule or a webhook event. Slack, GitHub and Linear are wired in. Backends, agent choices, automations and model settings live in the Agent Canvas frontend. A team can share one backend for code review while personal agents run on a laptop.
OpenClaw: one assistant that meets you in your chats
OpenClaw calls itself "an open-source AI assistant that runs on your own computer and meets you in the channels you already use" (OpenClaw). A single Gateway runs it, either as a personal assistant on a laptop or as a shared team deployment. Configuration is the only difference between the two. State, memory and credentials stay on your hardware. Models and harnesses are plugins you swap. There is no paid tier, hosted service or token.
The unit of work is a message. You write in Slack, Teams, Telegram, WhatsApp, Discord or iMessage, and the answer comes back in the same chat. For one person's assistant, or a small team sharing a channel, that shape fits well. The Gateway, the channels and the plugins make up the whole surface. There is no review step between the assistant's work and its effect, so a company that needs one builds it around the tool.
Hermes Agent: an agent that grows in the terminal
Nous Research calls its project "the self-improving AI agent". It builds skills from experience, searches its own past sessions, and runs from a full terminal interface or from Telegram, Discord, Slack, WhatsApp, Signal and the CLI. One gateway process serves all of it (Hermes Agent). hermes model switches models without code changes. A built-in cron scheduler delivers to any of those surfaces, and six terminal backends let it live on a $5 VPS or a GPU cluster.
This is the option for someone who works in a shell. The agent sits beside the commands, keeps notes between sessions, and answers in whichever chat app that person has open. Its configuration belongs to the gateway on the machine where it was installed. A team running agents on twelve machines keeps twelve setups in step.
Kortix: the rules are files, and a person merges
The configuration of a Kortix project is the repository itself. The agents, their skills, the company memory and the connectors are ordinary files in it, so what an agent may do sits in the repository rather than in a settings screen.
What a run may touch is written there. "Rule each tool call allow, ask or block", and the rule reaches the arguments a call was given, so a release note can pass while a refund waits. Models work the same way. "Kortix is model-agnostic. Pick the model per agent, per session or per message." The instance runs on "our cloud, your VPC, or your own on-prem network", which puts the same configuration in front of the security team that already reviews it.
The last step is the one the other three leave to the operator. A session clones the repo and works on a new branch. When the agent is done it commits and pushes that branch, and a reviewer reads the diff. "A person with the project.gitops.merge permission reviews the change request and merges it." An agent that rewrites its own prompt proposes that edit like any other, so a reviewer sees it beside the code the session touched.
The definitions behind these four tools are worth separating. What autonomous AI agents are turns the term into three testable parts. Autonomous vs human in the loop covers the three points where a person can hold a run. Running an agent with nobody watching follows a session from its trigger to its change request.
What buyers ask about open source autonomous agents
The shortlist is OpenHands, OpenClaw, Hermes Agent and Kortix. OpenHands runs coding agents from a self-hosted control center, OpenClaw is an assistant on your own computer, Hermes Agent works from a terminal, and Kortix treats agents, skills and memory as versioned files in the project's own repository.
Yes. All four run on hardware you control. Kortix runs on its cloud, your VPC or your on-prem network; OpenHands runs locally, in Docker, on a VM or in the cloud; OpenClaw runs on your own computer; Hermes Agent runs on a VPS or a GPU cluster.
Yes. Kortix is model-agnostic. Pick the model per agent, per session or per message, with your own provider key, a ChatGPT subscription or an OpenAI-compatible endpoint. OpenHands, OpenClaw and Hermes Agent each document bringing your own model too.
Kortix puts the agent's rules, memory and connectors in a repository your team can read, and lands each run as a change request someone reviews and merges. OpenHands runs coding agents from a control center, and OpenClaw is an assistant that lives in the chat apps you already use.
Put the rules in a repo, then add the trigger.
Start with the definition of autonomous agents, the three points where a person can stand, or a run that starts with nobody present.