Oh my, I recently read an article and then listened to the TED talk that the AI revolution is actually “underhyped.” And I couldn’t agree more. His reasoning is because of the potential of AI Agents.

In the future, Schmidt argues that AI agents will handle tasks by working together in various ways to handle complex business processes and solve critical problems. And it’s this process of multiple AI agents working together that we’re going to focus on in this post.

I first explored this concept while at Canon Medical, where we explored how multi-agent AI models can integrate patient care throughout the clinical pathway, creating a more cohesive healthcare experience from diagnosis to treatment and follow-up. This journey into the possibilities of collaborative AI is at the heart of our immersive customer experience at Origin Spice.

My Journey with Multi-Agent Systems

When I first started exploring AI solutions for Origin Spice, I honestly felt overwhelmed. We had multiple customer touchpoints that needed enhancement—from recipe recommendations to origin storytelling, freshness tracking to custom blend creation. It seemed impossible that one AI solution could effectively handle it all.

That’s when I stumbled into the world of multi-agent architectures. The concept immediately clicked for me: instead of one AI trying to be everything to everyone, why not create a team of specialized AI agents working together? Each could excel at specific tasks while contributing to a cohesive customer experience.

Models We Evaluated

Our team spent three months evaluating different multi-agent frameworks. Here’s what we discovered:

The Hierarchical Architecture initially appealed to us because of its transparent chain of command. Having a lead agent delegate to specialists seemed like it would create accountability and reduce confusion. However, in testing, we found it somewhat rigid when handling unexpected customer queries.

We then explored Collaborative Networks in which agents work as peers. I loved how flexible this was—agents could directly communicate without a middleman. However, coordinating consistent responses became challenging, especially when multiple agents had overlapping expertise.

The Marketplace Architecture was fascinating. It essentially lets specialized agents compete to handle specific customer requests. While theoretically efficient, it created latency issues as the system evaluated which agent should respond to each query.

After several months of testing, we ultimately created our hybrid approach that combines a lightweight hierarchy with shared memory and direct peer-to-peer communication when needed.

What Actually Worked For Us

Our winning approach was much more straightforward than I initially imagined. We created four specialized agents with clear domains:

• A Discovery Agent that helps customers explore new flavors
• A Story Agent focused on origin and cultural context
• A Recipe Agent specializing in cooking applications
• A Personalization Agent tracking preferences and making recommendations

Rather than building complex coordination mechanisms, we focused on two key principles:

1. Shared Memory: All agents access the same customer data and interaction history
2. Clear Handoffs: We designed specific triggers for when queries should transition between agents

The biggest surprise? The system worked best when each agent was trained to recognize its limitations and explicitly request help from other agents. Just like a good team of humans, our AI agents became more effective when they knew when to collaborate.

Results That Made It Worthwhile

I admit I was skeptical about whether customers would notice or care about our multi-agent approach. The results quickly changed my mind.

Average session times increased by 47%. Customer feedback mentioned “conversations that felt natural” and “information that connected.” Most importantly, our key metrics around product discovery and repeat purchases significantly improved.

The most satisfying moment came during a user interview when a customer said, “It’s like talking to a team of spice experts who all know me personally.” That’s exactly what we were aiming for.

How to Approach Your Selection Process

If you’re considering implementing multi-agent AI, here’s my practical advice based on what worked (and didn’t work) for us:

1. Start with customer needs, not agent architecture. Map out the different types of value you want to provide, then determine if these naturally cluster into distinct agent roles.
2. Prototype with simplified agents. We wasted time building complex systems before understanding the basic interaction patterns. Start with 2-3 simple agents and observe how they collaborate.
3. Focus on transitions. Most systems break down during the handoff between agents. Design these transitions carefully and test them extensively.
4. Create shared context. To maintain conversation coherence, ensure all agents can access the same customer history and product information.
5. Include human oversight. We still have human experts reviewing agent outputs, especially for creative recommendations or handling edge cases.

Looking Forward

As we continue evolving our multi-agent system, I’m most excited about increasing the agents’ specialization while maintaining a seamless customer experience. We’re also exploring ways to incorporate more community knowledge, allowing our system to learn from customer interactions in more sophisticated ways.

The multi-agent approach transformed not just our customer experience but how I think about AI implementation generally. Rather than seeking one perfect model, I now look for ways to combine specialized capabilities into collaborative systems.

If you’re embarking on your multi-agent journey, I’d love to hear about your experiences. What architectural approaches are you exploring? What challenges have you encountered? Share your thoughts in the comments below.