A continuation of the journey from isolated models to orchestrated intelligence
Key Takeaways:
Your AI team already exists Gemini, O3, Flash, Claude they just can't talk to each other
Zen MCP lets claude code talk to other LLM's
44 lines of code can unlock multi-model orchestration worth 10x productivity gains
The future isn't single AI assistants but orchestrated teams reaching consensus
Agentic architects build bridges between intelligences, not just systems
Tools matter less than mindset: Stop thinking AI as tool, start thinking AI as team
Welcome to the frontier.
If you're here from the LinkedIn post, you know the vision. Multiple AI models working as a team. Reaching consensus. Building better solutions together than any could alone.
Now let me show you how 44 lines of code made it real.
The difference between vision and reality is usually a small patch that everyone else overlooked.
The journey started with a realization. I had all the pieces but couldn't connect them.
Claude Code—Anthropic's coding assistant that can actually run terminal commands, spawn multiple editors, and execute autonomous workflows. "Agentic" means it can take actions, not just chat. The closest thing we have to a real AI developer.
Requesty—a unified API that lets you access Claude, Gemini, OpenAI, and Mistral with one key. No juggling multiple subscriptions. One interface, all the models.
Zen MCP—a Model Context Protocol server that promises true AI orchestration. As they describe it: "Claude Code for Claude Code." Not magic. Just super-glue. Claude stays in control, but YOU call the shots. You craft the prompts that make Claude bring in other models—or fly solo. You're the guide. The prompter. The puppeteer. You are the AI - Actually Intelligent.
Three powerful tools that refused to talk.
The error was mundane:
Error: Provider 'requesty' not recognized
Failed to initialize orchestration layerBehind every mundane error is an opportunity. This one was telling me that Requesty didn't know how to introduce itself to Zen.
Shanethacker had already started building the bridge with PR #52. He'd added Requesty as a provider. The foundation was there. But it couldn't speak to Zen MCP's orchestration layer yet.
In human terms: Imagine having Gemini, O3, and Claude in a conference room, but nobody knows each other's names. No introductions, no collaboration.
My patch completed the bridge. Built on shanethacker's foundation, it teaches Requesty to introduce its models to Zen MCP. Two methods, 44 lines:
// requesty.js patch
getModelList() {
const models = this.config.models || {};
const restrictions = this.config.restrictions || [];
return Object.keys(models)
.filter(model => !restrictions.includes(model))
.map(model => ({ id: model, name: models[model].name }));
}
getFullModelInfo() {
return this.config.models || {};
}The first method lists available models while respecting any restrictions you've set. Don't want to accidentally use expensive models? It honors that.
The second method is more comprehensive. It shows everything, including aliases. Because sometimes "gemini-pro" is also known as "google/gemini-2.5-pro". Small details that matter when you're building bridges between systems.
Code is just frozen thought. The best code thaws barriers between ideas.
As an Agentic architect, I see code as infrastructure for intelligence. These methods aren't just functions. They're translation layers.
Zen MCP speaks one language. Requesty speaks another. My patch teaches them to understand each other.
The beauty is in what it enables:
"Use Gemini to analyze this architecture"
"Ask O3 for optimization strategies"
"Get consensus from all models on this approach"No configuration gymnastics. No wrapper scripts. Just natural commands that orchestrate your AI team.
The implementation is deliberately simple. Complex systems fail. Simple bridges endure.
First, get Zen MCP (it's open source):
git clone https://github.com/BeehiveInnovations/zen-mcp-server
cd zen-mcp-server
npm installAdd your Requesty key to .env. Apply the patch (add those two methods to make Requesty talk to Zen). Your AI team is ready.
I remember the first orchestrated session. My n8n workflow prompts were taking 60 seconds to run. They worked perfectly - 100% success rate but something felt wrong.
"Use Zen to analyze these prompts and find consensus on optimization strategies."
Claude took the lead. Examined my 850+ lines of prompt engineering. Then brought in the specialists through Zen MCP.
The consensus was brutal and beautiful.
Mistral: "This is safety theater. You're over-engineering." Claude (via Zen): "60-70% of these constraints are unnecessary." Both agreed: "Trust the models more. Your Phase 1 needs just this:"
// Simplified prompt - 80% shorter
"Generate JSON for n8n workflow node"
// Core constraints only
"Must be valid JSON"
"Must include required fields"I asked: "But what about preventing JSON failures?"
The team consensus: Keep only what prevents parsing failures. Move validation to post-processing. Cut repetitive examples. Simplify positioning rules.
The result? 60 seconds down to 25 seconds. Same 100% success rate. Better natural output.
This is consensus. Not one model's opinion. Multiple perspectives converging on truth: I was protecting against problems that didn't exist.
In nature, evolution happens through competition and cooperation. In AI, innovation happens through orchestrated disagreement.
As an Agentic architect, I operate at the intersection of human intuition and AI capability. This patch embodies three principles:
Enable, Don't Force. Tools should unlock potential, not create dependencies. You can still use models individually. The patch just makes collaboration possible.
Connect, Don't Control. AI models need conversation space, not command structures. Zen MCP provides the room. My patch opened the door.
Guide, Don't Dictate. Human wisdom shapes AI consensus through gentle steering. Those "refine further" notes are more powerful than any prompt engineering. Because as Zen MCP reminds us: You are the AI - Actually Intelligent. The models are just waiting for your direction.
Let's talk tokens. Multi-model orchestration costs more. Significantly more.
Single model: $0.01-0.05 per complex query Orchestrated team: $0.10-0.30 per complex query
But consider what you get. A senior architect reviewing your code. A logic specialist catching edge cases. A performance expert optimizing algorithms. All debating until they reach consensus.
I pay my human senior engineers $200+ per hour. My AI team costs $0.30 per complex analysis. The ROI is obvious.
With Claude Max reducing base costs, the economics get even better. Use subscription for Claude's orchestration. Pay per-token only for specialist consultations.
The best investment isn't in owning resources. It's in connecting them.
Silicon Valley built walled gardens around each model. Anthropic has Claude. Google has Gemini. OpenAI has GPT. Each promises to be your everything.
But innovation happens at the edges. In the spaces between systems. Where different perspectives collide.
The vendors won't build these bridges. There's no incentive to help competitors. That's why it falls to us. The builders. The architects. The ones who see potential where others see barriers.
Shanethacker saw the gap and started building. I found his work and completed the connection. My 44-line patch is built on his foundation. That's how the community works. We build on each other's bridges.
There are thousands more waiting to be built.
If you're a founder, you have an AI team waiting. They're just in separate rooms.
If you're a developer, you can build bridges. Look for the gaps. Find the incompatibilities. Write the patches.
If you're an investor, watch for Agentic architects. We're the ones connecting systems that vendors keep apart.
The tools exist. Claude Code for agentic workflows. Zen MCP for orchestration protocol. Your choice of API. Simple patches to connect them.
The missing piece is mindset. Stop thinking about AI as a tool. Start thinking about AI as a team.
Your architect: Gemini Pro Your debugger: O3 Your implementer: Flash Your coordinator: Claude Your leader: You
The future belongs to those who can conduct the orchestra, not those who own the instruments.
We're moving from single AI assistants to AI teams. From isolated intelligence to orchestrated consensus.
The technical barriers are trivial. 44 lines proved that. The mental barriers are what hold us back.
Every time you prompt a single model, you're using a fraction of available intelligence. Every time you accept the first answer, you're missing the debate that leads to better solutions.
Start small. Use Zen MCP to get a second opinion on your next technical decision. Watch how models challenge each other. Notice how consensus emerges.
Then scale up. Architecture reviews with multiple perspectives. Debugging sessions with specialized models. Implementation with consensus-driven design.
This is Agentic architecture. We don't just build systems. We build the bridges between them. We create the conditions for intelligence to flourish.
The patch is public. The tools are free. Your AI team is waiting.
Consensus isn't agreement. It's the wisdom that emerges when different intelligences collide, debate, and refine.
Welcome to the frontier. Welcome to consensus. Welcome to being the Actually Intelligent one in the room.
I'm an Agentic architect specializing in bridging human intuition and AI collaboration. The Requesty implementation started with shanethacker's PR #52. My patch completes the integration for Zen MCP orchestration. For implementation help and orchestration strategies, let's connect.
The future isn't one AI. It's AI teams reaching consensus. Are you ready to lead yours?