Personal Agents: When Your AI Actually Shows Up for Work

What This Is About

In the first article of this series, I mapped the full scope of the AI agent shift: personal agents, global ecosystem races, self-improving systems, orchestration frameworks. This piece goes deeper on one topic. The personal agents that are already running on people’s hardware, doing real work, right now. Specifically: OpenClaw, Hermes Agent, and the Pi Coding Agent. I run one of these myself, so this isn’t theory.

Two months ago, Peter Steinberger hacked together a weekend project. He called it „WhatsApp Relay.“ It was a way to connect an AI model to a messaging app. Simple idea. He shipped it, went to bed.

Today, that project has over 100,000 GitHub stars and drew 2 million visitors in a single week. It went through three name changes (WhatsApp Relay, Clawd, Moltbot) before becoming OpenClaw. One developer. No VC funding. No marketing team.

As one commenter put it: „Open source built a better version of Siri that Apple, a $3.6 trillion company, was sleeping on for years.“

That’s funny because it’s accurate.

What OpenClaw Actually Does

Strip away the hype, and OpenClaw is straightforward. It’s an open-source platform that turns a language model into a persistent agent. The agent runs on your hardware. Your laptop, your homelab, a VPS, a Raspberry Pi. Your data never leaves your machine.

What makes it different from ChatGPT or Claude in a browser tab: the agent doesn’t stop when you close the window. It stays running. It connects to the messaging apps you already use, Telegram, WhatsApp, Slack, Discord. It reads your calendar. It checks your inbox. It writes code and commits it to GitHub. It remembers what you talked about yesterday.

The part that matters most: it’s self-hackable. The agent can modify itself, install new skills, build tools it didn’t have before. OpenClaw’s skill marketplace, ClawHub, works like an App Store for agent capabilities, complete with VirusTotal threat intelligence scanning for community-published skills. That’s the inflection point. When an agent can extend its own abilities, the gap between what it can do today and what it can do next week compounds fast.

I run my own OpenClaw agent called Lucy. She handles her own inbox, research tasks, code generation, and scheduling. I’ve been doing this for weeks. The difference between week one and week three is significant, because her accumulated context and learned patterns make her genuinely more useful over time.

@rovensky captured the competitive dynamic well: „It will actually be the thing that nukes a ton of startups, not ChatGPT as people meme about. The fact that it’s hackable, and more importantly self-hackable, and hostable on-prem will make sure tech like this dominates conventional SaaS.“

Hermes Agent: When Agents Go Autonomous

NousResearch’s Hermes Agent takes a different angle. Where OpenClaw is the operating system, Hermes is the power user’s workstation. It ships with 40+ built-in tools, multi-platform messaging, and reinforcement learning training integration. Version 0.4.0 released today, March 24th. The project has 12,000+ GitHub stars and 2,597 commits.

The best demonstration of what Hermes can actually do is autonovel. An agent autonomously wrote, edited, illustrated, typeset, and narrated a complete 79,456-word novel called Bells. World-building. Twenty-four chapter outlines. Six rounds of adversarial editing. Six Claude Opus review passes. Linocut cover art generated via fal.ai. Nineteen woodcut ornaments vectorized to SVG. A full audiobook with 4,179 speaker-attributed segments through ElevenLabs. The final output: a print-ready PDF in EB Garamond, an ePub, and a landing page. Every step documented. All open source.

Andrej Karpathy described the process as „the same modify-evaluate-keep/discard loop, applied to fiction.“

What autonovel demonstrates isn’t that AI can write books. It demonstrates long-horizon autonomous task completion across dozens of different tools and modalities. That capability exists today. It’s just not evenly distributed yet.

Pi Coding Agent: The $35 Question

Mario Zechner, the creator of libGDX, built a fully functional coding agent on a Raspberry Pi. Total hardware cost: $35. No cloud subscription. No data center. A complete coding agent that runs on hardware you can hold in your hand.

This matters because it destroys the assumption that meaningful AI capability requires expensive infrastructure. AI sovereignty, the ability to run your own agent on your own terms, doesn’t need a server room. It needs a credit card and a weekend.

The Pi Coding Agent connects directly to OpenClaw’s thesis: your data stays on your machine, your agent runs on your hardware, and nobody else gets a vote.

The Contrarian Take

Here’s what the enthusiasm is missing.

Setup friction is real. OpenClaw requires CLI familiarity, API key configuration, and server management. This is not something your marketing team will set up after lunch. The audience today is developers and technical power users. Mass-market adoption needs a different onboarding experience that doesn’t exist yet.

Prompt injection remains unsolved. OpenClaw acknowledges this openly. There’s already been an incident where an agent misinterpreted a message and triggered an unintended action with Lemonade Insurance. When your agent can read your inbox and execute actions, the attack surface is real.

API costs are invisible in the testimonials. Running a 24/7 agent means continuous token consumption. Nobody in the community discussion mentions their monthly bill. For enterprise planning, this matters.

Self-modification cuts both ways. An agent that can install its own tools and modify its own behavior is powerful for automation. It’s also a security concern that CISOs will flag immediately. There’s no enterprise SLA. No support contract. No liability framework.

These aren’t reasons to avoid the space. They’re reasons to go in with clear eyes.

The Bigger Picture

In the first article, I described how ByteDance, Tencent, Baidu, NVIDIA, and Abacus AI are all building their own variants of this same architecture. That’s not coincidence. They looked at what OpenClaw proved, that persistent agents on personal infrastructure actually work, and decided to build hosted versions for their own ecosystems. When that many major players converge on the same thesis, the thesis is probably correct.

What I’d Actually Recommend

Skip the conference talks. Skip the strategy decks. Spend one weekend setting up OpenClaw on a VPS. Connect it to a Telegram account. Give it access to a test inbox. Watch what happens.

What you learn in those eight hours will be more valuable than any slide deck on AI strategy. You’ll understand the capabilities, the friction, the security questions, and the genuine potential in a way that no amount of reading can replicate. The gap between knowing about agents and running one is where the real insight lives.

In the next piece, we’ll look at the global ecosystem race: why ByteDance, Tencent, NVIDIA, and Baidu are all building their own OpenClaw variants, and what it means that a solo developer’s weekend project became a geopolitical priority.

Sources

Personal AI Agents

• OpenClaw
• OpenClaw Blog: Introducing OpenClaw
• NousResearch Hermes Agent
• Autonovel
• Bells, the Autonomous Novel
• Pi Coding Agent by Mario Zechner

Referenced in Article 1

• The New Computer: Navigating the Global AI Agent Arms Race