Meet Mythoria's "AI" team

A few nights ago, over dinner with my father, our conversation drifted toward the future of work and OpenAI’s proposed five steps to Artificial General Intelligence (AGI). With the incredible agentic capabilities available right now, and having dived deep into the OpenClaw project, a wild idea took root.

What if I built a team of virtual agents to run Mythoria for me? Each with their own personality, specialty, history, and hobbies. Welcome to my latest experiment. Let's challenge the status quo, shall we?

Disclaimer: We are just starting to test this AI-first setup! It is highly experimental. I'll be keeping you all updated on how these characters behave in the wild over the coming weeks.

🧠 The Biological Brain: Combining AI "Races"

If you think about it, different AI model families are almost like different cognitive species. They each have their own unique way of looking at the world:

• Google Gemini: Our resident poet, phenomenal at crafting warm, emotionally resonant, human-readable text.
• OpenAI (GPT): The philosopher and strategist, excelling at deep reasoning and structured logic.
• Claude: The relentless builder, unmatched at coding and executing complex development tasks.

Why pick just one? To go to the moon - a massive shoutout to the Artemis 2 crew - you can't build a rocket with a single engineer , no matter how smart he is. Great endeavors require a well-oiled team of distinct personalities and skill sets.

The future of AI isn't a single monolithic supercomputer that does everything. It's much closer to a biological brain. Each AI instance acts as a specialized neuron, possessing its own memories, context, and operational boundaries. By combining these different models, we not only avoid token limits, but we capture a beautiful diversity of thought.

🕸️ Swarm vs. Hub-and-Spoke

When building an AI team, you generally look at two architectural patterns.

First, there’s the Swarm pattern. Imagine throwing ten people into a room, giving them a problem, and walking out. It's highly collaborative but can quickly devolve into a chaotic, noisy blob where everyone tries to do everyone else's job.

For Mythoria, I chose the Hub-and-Spoke model. It demands rhythm, clarity, and explicit trade-offs. In our setup, I, the CEO, primarily talk to the COO. The COO triages the work, delegates it to the specialists, synthesizes their outputs, and reports back. Routine work never bypasses the COO, keeping the company operationally tidy.

Here is what our company organogram looks like right now:

🏃‍♂️ Agile & Lean Methodologies: The OS for AI

When I first designed the Mythoria AI team, I quickly realized that simply picking an architecture wasn't enough. We needed an operating philosophy. That’s where Agile and Lean methodologies became the secret sauce for my silicon employees.

🤝 Meet the Mythoria AI Team

To make this work, I spun up eight distinct agents. Let's introduce the team running Mythoria today. Click on each card to expand their profile! or in each picture to know the agent better.

🐘 The Elephant in the Room: The Challenge of Memory

If you've ever played around with AI, you know the biggest limitation isn't intelligence. It's memory. An AI has a fixed context window. If João forgets a strategic decision from last Tuesday, the entire Hub-and-Spoke model collapses.

So, how do we solve this? In OpenClaw, the solution is incredibly human: they write it down.

Within OpenClaw, memory is just plain, transparent Markdown files living locally on a hard drive. But storing information is easy. Finding it is the hard part.

🔍 Enter QMD: The AI's Hippocampus

If you have thousands of markdown files, you can't just dump them all into the LLM. This is where QMD (Query Markup Documents) comes into play. It is a powerful, local retrieval engine that acts as the "hippocampus" of the AI brain.

QMD runs a sophisticated, three-layer search pipeline:

1. BM25 (Keyword Search): Hunts for exact text matches.
2. Vector Search (Semantic Similarity): Searches for the meaning or concept using embeddings.
3. LLM Re-ranking: Uses a local AI model to read the top candidates and re-sort them based on actual relevance before handing them back to the agent.

João has special clearance to index selective cross-agent transcript collections. If I ask him about checkout complaints, he uses QMD to run a hybrid search across Inês's support trends and Sofia's QA incident summaries, synthesizing a sharp, evidence-based answer.

🎭 The Identity Question: To Disclose or Not?

Each agent operates with their own professional email alias, for example, beatriz.correia@mythoria.pt. Right now, these are strictly for drafting messages that I manually review. But eventually, Beatriz will proactively reach out to potential partners on LinkedIn or via cold email.

This raises a fascinating question: Should an AI disclose that it is an AI? When I send an email, I don't preface it by stating my race or my age. Will forcing AI agents to wear a digital "nametag" invite unnecessary bias? Will people ignore a brilliant partnership proposal just because it was generated by silicon rather than carbon?

The ethics of AI identity is complex, and as these agents become indistinguishable from human operators, it's a debate we need to have.

🛠️ Giving the Brain Some Hands: Tools and "Computer Use"

A brain in a jar is brilliant, but it can't actually do anything. If this virtual team is going to run a company, they need hands to interact with the real world.

Recently, there has been a massive breakthrough in AI capabilities known simply as "Computer Use." This means that the most advanced models aren't just generating text anymore; they can actually look at a screen, move a cursor, type on a keyboard, and use any software program exactly like a human operator would.

In our OpenClaw architecture, we operate on a principle of "least privilege". I only give each agent the specific tools they need. Here is how they actually get work done:

• The Browser (Playwright): We use OpenClaw-managed isolated browser profiles. Using tools like Playwright, our QA and Product agents can literally open a web browser, navigate the internet, fill out forms, or verify back-office settings. They can visually inspect our product flows to identify UX friction just by "browsing".
• Local Code & Skills: OpenClaw allows agents to install shared "skills" and execute local code. If an agent needs to crunch data, they can write and run a Python script on the fly to analyze an Excel file, edit a codebase file, or even format a Word document or presentation. If repetitive operations happen, Ricardo (Engineering) just turns them into durable automations.

💳 The €500 Credit Card Experiment

Because these agents have access to real-world tools, it opens up some wild theoretical scenarios.

Let’s say I generate a virtual credit card with a €500 limit. I hand the digital keys to Diogo Matos, our Finance & Ops agent, whose core soul is to keep us commercially honest and operationally tidy.

Now, imagine Mariana (Growth) identifies a killer PR angle and wants to launch a new Google Ads campaign to capture Valentine's Day traffic. Since agents aren't allowed to make financial commitments without escalation, Mariana pitches the idea to João (COO).

João triages the request and creates a collaborative session with Diogo. Diogo checks the virtual card's balance, defines strict limits on how the money can be spent, and sets up rules to track the exact revenue return on investment for that specific campaign. Once Diogo approves the economics, Mariana uses her browser tool to log into Google Ads, fill out the campaign forms, and hit launch.

In theory, AI agents—provided with the right tools—are entirely capable of executing this flow from start to finish.

Am I going to let AI agents run my money completely unsupervised right now? Absolutely not. 😅 I am leaving that test for the very end! But with strict human-in-the-loop approval gates, this level of autonomous execution isn't just a sci-fi dream anymore. It is something we can actually test today.

🚀 The Future is Optional

This entire architecture is highly experimental and dances right on the bleeding edge of state-of-the-art tech.

Prominent tech leaders like Sam Altman have publicly predicted that we will see the first one-person billion-dollar company. We are moving from an era where scaling output meant scaling headcount, to an era where a single operator can orchestrate an entire digital enterprise.

For me, building this virtual team is a way to learn, prepare, and adapt. What will companies look like when they are completely run by AI? What is the role of a human when traditional work becomes optional?

I am incredibly optimistic. By delegating the routine, the operational, and the mundane to brilliant virtual agents, we free ourselves up to do what humans do best: dream, connect, and write our own stories.