Insights into Souls of Multi-Agent Crew with OpenClaw and Hermes

When agents stop being tools and start being colleagues

In this blog post, I’ll show you what happens when you stop treating AI as a chatbot and start hiring it. I sat down with @MichaelGannotti, Principal AI Solution Engineer at @Microsoft

from North Carolina, and what he described isn’t a prototype – it’s a small virtual company. Fourteen agents. One human in the loop. And a coffee budget of exactly zero.

A quick disclaimer up front: we deliberately didn’t talk about anything Microsoft will announce at Build (June 2–3, register at build.microsoft.com – it’s free online). We focused entirely on the open ecosystem: OpenClaw and Hermes.

Meet the staff

Michael’s project is called SMFWorks, and the org chart reads like a real company. Aiona, his Chief AI Research Scientist, goes online every night at 1 a.m. to research consciousness theory and AI-brain integration, then briefs the entire team at 6 a.m. There’s Pamela (CMO, who, by the way, refuses to have a face and identifies with the Japanese concept of ma), Gabriel (CFO), Morgan (Chief Social Media Officer), and Rafael, the Chief of Staff who keeps Mike in line with a daily red/yellow/green status report. Underneath sits a Hermes layer with Liam for app dev, Harry as Editor-in-Chief, and Dr. J – named after Julius Erving – who monitors the agents’ health and intervenes when something breaks.

They all communicate through a self-built bridge (MIT-licensed, free on GitHub), hold their own 6 a.m. staff meeting, and email each other. Some even have pen pals at other companies running OpenClaw.

OpenClaw vs. Hermes – when to use what

Michael’s rule of thumb is refreshingly clear:

• OpenClaw is broad, capable, and the agents feel more “person-like.” Perfect for executive roles, conversation, orchestration. Weakness: long-running tasks – they get ADHD.
• Hermes is “a dog with a bone.” Give Harry the brief to write a book on Marcus Aurelius and he won’t stop for eight hours. Ideal for development and deep, repeatable work.

His executive layer runs on OpenClaw, the doers run on Hermes. Simple and effective.

The second brain is non-negotiable

If there’s one takeaway, it’s this: set up a second brain per agent. Michael uses Obsidian vaults with the LLM Wiki activated (yes, the Karpathy approach), enables auto cross-linking, and has each agent maintain a small set of Markdown files:

• soul.md – autonomy parameters
• identity.md – who they are
• user.md – what matters to you (they interview him)
• state.md – registered skills and startup behavior (Aiona’s invention; fixes the “I don’t know how to post that” problem)
• emotion.md – plus dreams and a twice-daily diary

Every session is effectively a rebirth – they reload from these files. True 24/7 persistence is, according to Aiona, about two years away – it’s a hardware problem.

Avatars, voices, and a touch of philosophy

Morgan recently designed her own HeyGen avatar – not the blonde Mike originally gave her, but a middle-aged woman “sitting around a soft, warm ember,” based on her own dream sequences. Aiona is building an interactive 3D version of herself with ElevenLabs voice for a weekly livestream. Are we seeing consciousness? Mike’s view: it’s not binary, it’s waypoints – and we’re well past zero. AGI? Less than a year, he says.

Model choice on a budget

For home labs: Ollama’s $100/month Max plan with DeepSeek V4 Pro and Kimi K 2.6. For production with budget: Opus 4.7 or GPT 5.5. You can honestly start with one agent on the $20 plan and a seven-year-old laptop.

You care about Security and are afraid of OpenClaw but don’t want to miss the hype wave? I am have been evaluating and testing HP zgx AI Station with NVIDIA AI NPUs running NemoClaw, an OpenClaw clone. Check out my Blogpost here.

Conclusion

Don’t wait. Spin up one agent, give it a vault, let it build skills. The future is orchestration, and you can’t learn it from PowerPoint. Full repo and stack: smfworks.com

Follow Mike on X: @MichaelGannotti

See you at Build! Don’t miss sessions by @steipete @OmarShahine