Alex Finn was the OpenClaw guy.
Strangers recognized him at poker tournaments for it. His whole feed was OpenClaw setups, OpenClaw routines, OpenClaw everything.
Last week he sat down with Greg Isenberg and said he's done. He moved everything to Hermes Agent Desktop, and he walked Greg through the entire setup on camera: how he cut agent bills that hit $1,000 a month for some users, how he runs 7 different agent profiles, and the cron job that scans Reddit and X every 20 minutes and builds him app prototypes while he sleeps.
I watched the full walkthrough so you don't have to. Here's the playbook, with the parts that actually matter for a bootstrapped founder.
Why the OpenClaw guy switched
Alex compared his move to Michael Jordan leaving basketball for baseball in 1994. Half a joke. But he built a real audience on OpenClaw content, and he walked away from those easy clicks anyway.
His reasoning: the two teams went in opposite directions. Hermes shipped focused weekly updates, each one polished and tested. OpenClaw went shotgun mode, cramming in power-user features that broke his install more than once.
"It doesn't need to be a CLI," is how Greg summed it up. Alex gets constant comments from viewers who see "just paste this into your terminal" and immediately close the video. A desktop app removes that wall.
If you're new to this whole space, start with my breakdown of Greg's earlier episode on what an AI agent loop actually is. This article assumes you know roughly what an agent does.
Sessions: the fix for $1,000 a month agent bills
The number one complaint Alex hears about Hermes: "I'm paying $1,000 a month."
The cause is almost always one giant thread.
Most people talk to their agent in a single mono-thread. Stock research, content ideas, coding, random questions, all in one conversation. Every new message re-sends that entire history to the model. That alone can 3x or 4x your bill.
The desktop app fixes this with sessions. Every conversation gets its own thread, you can pin them, and you can sort them into folders. Alex keeps a content session, a stock research session, a local model session. Each one stays slim, so each message costs a fraction of what the mono-thread version would.
→ One topic, one session. That single habit is most of the savings.
Profiles: one agent per model, not one per job title
Alex runs 7 profiles: Hermes, Billy, Coder, GPT-Me's, Librarian, Oracle, and Qwen. Each profile is a fully separate agent with its own skills, personality file, and memories.
Here's where his take gets contrarian. Most people build profiles like an org chart: a product manager agent, a designer agent, a CTO agent. Alex thinks that's a waste of money. When your "CEO" agent delegates a thumbnail to the "product manager" who delegates to the "designer," every hop burns context and cost. You could have just asked one smart model directly.
He organizes by model instead:
- Opus 48 (his default) for high-level thinking, planning, and orchestration. Smartest model he's used, also the priciest.
- GPT 5.5 for coding. He thinks it codes better than Opus right now, and the usage limits are far more generous.
- Qwen 37, running locally on the DGX Spark sitting on his desk. Free unlimited intelligence for research and background tasks.
Know each model's strengths, route the task to the cheapest model that can handle it. That's the whole system.
And the difference between profiles and sub-agents, since everyone asks: sub-agents are copies of the same agent for parallel work (build 6 features of a micro SaaS at once). Profiles are for genuinely different skill sets. Different skills, new profile. Same skill, many threads, sub-agents.
The features most people miss
Three things from the walkthrough that Alex called "sneaky" good.
Artifacts. Every link, image, and file you've ever sent your agent, auto-organized and searchable in one place. Alex used to maintain a manual second-brain system where he'd tell his Librarian agent "file this under X." Now he just drops links with no instructions and finds them later by searching. The product team clearly watched how people were hacking together second brains and just built it in.
The skills panel. Your agent ships with over 150 skills installed, and every single one adds context to every message. That's silent cost. The desktop app lets you toggle off what you don't use. It also shows you the skills your agent has been quietly creating in the background based on your habits, which most users never realize is happening.
The cron section. Scheduled tasks are where agents turn into actual business automations instead of a fancy chatbot, and this was the most broken part of the old workflow. You'd tell your agent "do this every morning at 9am," it would say sure, and nothing would happen. No way to verify. Now there's a panel listing every scheduled job, and you can create them manually with full confidence they exist.
The brain dump + reverse prompt trick
Best practical tip in the episode, and it takes 60 seconds.
When Greg asked how to write cron jobs that actually work, Alex demoed his formula live. Most people write: "Set up a morning brief about AI news." Then they wonder why the output is stale headlines in a wall of text.
His version has two steps:
- Brain dump. Tell the agent who you are. "My interests are AI, stock investing, tech, and the Boston Celtics."
- Reverse prompt. Ask: "What would be the best prompt I can use to set this up with you?"
The agent wrote him a prompt he'd never have written himself: pull only from the last 24 hours, don't rely on memory, use real headlines and prices, cover these 4 areas, lead with a one-line summary, format with bullets.
You're talking to something smarter than you at prompt-writing. Let it write the prompt. (Why would you wing it?)
The use case that matters: an agent that finds problems to solve
This is the part worth stealing.
Alex has a cron job called "Daily AI business opportunity scan." Every 20 minutes, his local Qwen model reads Reddit and X looking for people complaining about real problems. It compiles everything into a custom dashboard that shows:
- The challenge it found, with a link to the source thread
- Why Alex specifically is positioned to solve it (the agent knows his skills, his audience, his existing products like Vibe Coding Academy)
- A suggested first move
And when the agent decides an idea is strong enough, it goes one step further and builds a working micro SaaS prototype on its own. There's literally a button on the dashboard: "Henry built a prototype." Click it, test the idea, kill it or ship it.
He admits the dashboard looks like AI slop. He doesn't care. It's an employee that works around the clock finding him markets.
One caveat he flagged: every 20 minutes only makes sense because his local model is free. On Opus or ChatGPT you'd be "paying out the wazoo." Run it once a morning and you get 90% of the value.
The deeper lesson is his framing of solopreneurship: find other people's challenges and solve them. As a one-person company with no overhead, you can profitably serve tiny markets that no funded startup would touch. An agent that hunts those markets for you 24/7 is pure leverage.
Do you need a $4,800 box on your desk?
No. Get that out of the way first.
Alex runs his free background tasks on an Nvidia DGX Spark: $4,800 (recently raised from $4K), 128GB of unified memory, plug it into the wall and it works. He'd actually prefer a Mac Studio for the nicer interface, but the good configs are sold out everywhere. His bet is that hardware and memory prices keep climbing through the AI build-out, so buying now beats buying later.
Greg's take was more grounded, and it's the one I'd give you: download the desktop app first, prove to yourself that agents create real value in your week, and only then think about hardware. Don't be the person with a $5K box collecting dust next to the Peloton.
Alex's reframe on cost is still worth sitting with. People compare a $200/month agent plan to Netflix and call it expensive. Wrong category. Netflix takes your money and gives you nothing back. An agent, used right, is an employee. If it makes you 10x the cost, the price was never the question.
If you want to go deeper on making agents part of your actual daily workflow, I broke down Greg's full 60-minute method in how to become AI native.
FAQ
What is Hermes Agent Desktop?
A desktop app for running Hermes AI agents without touching a terminal or Telegram. It adds visual session management, agent profiles, a searchable artifacts library, a skills panel, and a cron job manager for scheduled tasks. Alex Finn calls it the best AI agent experience available right now.
How do I stop my AI agent from costing so much?
Three moves from the episode. Split your conversations into separate sessions so you stop re-sending a giant history with every message. Route tasks to the cheapest capable model (expensive model for planning, cheaper ones for coding and research). Turn off the preinstalled skills you don't use, since each one adds context cost to every message.
What's the difference between a profile and a sub-agent?
A profile is a fully separate agent with its own skills, personality, and memories. A sub-agent is a temporary copy of your current agent, spun up for parallel work. Need different skill sets, create a profile. Need the same skill applied to 6 tasks at once, let it spawn sub-agents.
Do I need a local model to copy the opportunity scanner?
No. The local model just makes it free to run every 20 minutes. On a cloud model, schedule the same scan once a day. You still get a daily list of real problems people are complaining about, matched to your skills.
Steal playbooks like this every week
This is exactly the kind of episode I break down on the Profitable Founder Podcast: real operators showing the actual systems behind their numbers, not theory.
Every week I sit down with bootstrapped SaaS founders doing $100K to $10M a year and get them to open the hood.
