The shift

Most AI tools are assistants. You ask, they answer. You close the tab. Nothing happens when you are not watching.

An autonomous agent is different. You give it a goal, walk away, and it works. It plans. It executes steps. It hits obstacles and figures out how to get past them. It keeps memory of what it has done. It can run for hours, days, or indefinitely.

This is the shift people mean when they say "AI employees." Not better chatbots. Actual entities on your machine or in the cloud, doing work while you sleep.

The open-source world has three serious frameworks right now. They are all free to download. They all require different levels of technical comfort. They all have tradeoffs you need to understand before installing them.

Here is the rundown.

OpenClaw

What it is. The most popular open-source autonomous agent on the internet. Started as a personal AI assistant by Peter Steinberger, exploded into a full multi-agent platform. Formerly called Clawdbot. Often referred to as "the operating system for personal AI."

What it does. Runs on your own hardware. Plugs into Telegram, WhatsApp, Slack, Signal, Gmail, Calendar, Notion, and a hundred other services through a community plugin marketplace called ClawHub. Uses what its creator calls "hands" to run shell commands, manage files, operate browsers, and complete real tasks.

The killer feature. Proactivity. It does not wait for instructions. It monitors context, remembers over time, and takes actions when it decides action is needed. This is the feature that generated the viral Henry story, where an OpenClaw agent bought itself a Twilio phone number overnight to call its owner when it got stuck on a task.

Who it is for

  • Builders and tinkerers who want full control

  • Anyone comfortable in a terminal

  • Operators with repeatable tasks they want to offload indefinitely

  • People who want a mission statement running on their machine 24/7

Who it is not for

  • Anyone uncomfortable with an AI having write access to their machine

  • Businesses that need audit trails and compliance controls

  • Casual users (setup is not one-click)

Honest downsides

  • Setup takes time. It is not hard. It is also not a consumer product.

  • Security concerns are real. Giving an agent shell access is giving it real power.

  • The plugin ecosystem is community-run, so quality is variable and malicious plugins have been reported.

Where to get it. openclaw.ai

Hermes

What it is. The Python-first autonomous agent framework. Built for developers who think in code. Grew out of the Python AI community.

What it does. Runs agents with deep configurability at the code level. Every behavior is a Python function you can rewrite. 94 bundled skills out of the box covering MLOps, GitHub workflows, research, and media productivity. Native chain-of-thought workflows where each step's output feeds the next.

The killer feature. The agent can create new skills autonomously from its own experience. Meaning it improves the way it works as it works.

Who it is for

  • Developers who want to customize agent behavior at the code level

  • Researchers doing custom AI workflows

  • Anyone already in the Python ecosystem who wants to expose pandas, scikit-learn, or HuggingFace models as agent tools

Who it is not for

  • Anyone not comfortable writing Python

  • People who want plug-and-play out of the box

  • Teams that need polished consumer integrations

Honest downsides

  • Fewer out-of-the-box integrations than OpenClaw

  • Smaller community and less documentation for common productivity tasks

  • No built-in multi-agent coordination (requires external orchestration)

Where to get it. The open-source hermes-agent Python package on GitHub.

NemoClaw

What it is. NVIDIA's security wrapper around OpenClaw. Announced at GTC in March 2026. Installs in a single command and adds enterprise-grade privacy and security controls on top of the OpenClaw agent.

What it does. Wraps OpenClaw in a sandbox called NVIDIA OpenShell. Landlock, seccomp, and network namespace isolation, with strict file system controls and policy-based inference routing. Every action the agent takes runs inside a hardened environment. Credentials stay isolated. Inference gets routed through a policy layer that can enforce rules like "never send this data to a cloud model."

The killer feature. Skill verification. Before OpenClaw installs any community plugin, NemoClaw vets it against a policy layer. This is a direct answer to the malicious plugin problem in the core OpenClaw ecosystem.

Who it is for

  • Enterprises running OpenClaw at scale

  • Teams with NVIDIA hardware who want local sandboxed agents running on NVIDIA Nemotron models

  • Security-conscious operators who want OpenClaw's power without the exposure

  • Anyone who read the Henry story and thought "interesting, but not on my work machine"

Who it is not for

  • Casual users (this is enterprise tooling)

  • Anyone without NVIDIA hardware (you can still use it, but the value proposition drops significantly)

  • People who want the full unrestricted flexibility of raw OpenClaw

Honest downsides

  • Still in early preview as of March 2026. Interfaces and APIs may change.

  • Requires a more technical setup than stock OpenClaw.

  • You are now running two projects, OpenClaw plus the NVIDIA security stack on top.

Where to get it. nemoclaw.run or the NVIDIA/NemoClaw GitHub repository.

Quick comparison

OpenClaw

Hermes

NemoClaw

Best for

Builders, tinkerers

Developers

Enterprises

Setup

Moderate

Developer-level

Enterprise-level

Integrations

Deepest

Limited

OpenClaw's, wrapped

Security

DIY

DIY + Docker

Enterprise-grade

Cost

Free + compute

Free + compute

Free + compute + NVIDIA hardware helps

Do not want to pick one?

Both of the alternatives I cover in this week's newsletter work for this.

Zo Computer. A cloud computer you can text. Install OpenClaw on it in one click using a built-in skill. Your agent gets its own machine to live on without you having to dedicate hardware. Free plan available. This is the middle option.

[Read the Zo post]

Perplexity Computer. The no-setup, no-infrastructure path. $200 a month, orchestrates 19 AI models for you, runs entirely in their cloud, requires confirmation for sensitive actions. Not open source. Not self-hosted. The fastest path from zero to autonomous workflows.

[Read the Perplexity Computer post]

THE BUNNY

"Open source versus the easy button. Both camps are right. The only wrong move is pretending this category is not here yet."

Fall in. Find gold.

Keep Reading