In this exclusive interview, we sit down with Mathis Joffre, an engineering leader specializing in AI infrastructure. Mathis has a distinguished background, having contributed to AI platform development at one of Europe’s largest cloud providers.
Now, he’s focused on the cutting edge of runtime design, model orchestration, and developer tooling for autonomous systems, driven by a passion for making advanced research deployable in production at scale. Join us as we explore his journey and insights into the rapidly evolving world of AI agents.
You’ve worked on AI infrastructure both at large cloud providers and now in building next-gen systems. What first drew you to this space, and how did that experience shape your view of how AI agents should run today?
I got into AI infrastructure while working at OVHcloud, where I helped design distributed systems for millions of users. That experience revealed the limits of traditional cloud models, especially for emerging workloads like AI agents, which need fast, isolated, and resumable execution.
Curious to explore better solutions, I focused on microVMs and edge computing, sharing what I learned through talks and open source. That journey led me to co-found Blaxel, where we are building infrastructure specifically for AI agents, optimized for tool use, low latency, and secure multi-region execution.
Looking back, my experience in both enterprise cloud and startup R&D shaped a clear belief. AI agents need infrastructure that is lightweight, reactive, and agent-aware. They require more than just scaled-down versions of legacy cloud systems.
For those unfamiliar, what exactly are AI agents, and how are they different from traditional AI tools like chatbots or recommendation engines?
AI agents are autonomous systems that can perceive, reason, and act across multiple steps to achieve a goal. Unlike traditional AI tools like chatbots or recommendation engines, which are usually reactive and single-purpose, agents are proactive and task-driven.
Think of a dental receptionist. A chatbot might simply list available time slots. An AI agent could answer your call, hold a conversation, respond to follow-up questions, update your records, and book your appointment, all without human intervention.
What makes AI agents distinct is their ability to plan, use external tools and APIs, maintain context, and adapt over time. They often work in coordination with other systems or agents to complete tasks.
You’ve said that where these agents “live,” on the edge or in the cloud, matters more than people think. What’s the core trade-off you’ve seen between edge computing and centralized infrastructure?
The core trade-off is between control and capability.
The cloud offers scale, compute power, and centralized coordination, which is ideal for training models or managing complex workflows. The edge provides lower latency, improved privacy, and greater resilience. This is especially important when agents are interacting in real time or need to operate locally, such as on a mobile device or in an industrial setting.
The best systems often combine both, using centralized intelligence together with local autonomy.
In your time building platforms to support AI agents, what real-world challenges have made centralized cloud compute feel limiting? Can you give an example that surprised you?
One of the biggest issues we encountered was latency, particularly in tool-use scenarios. Agents making multiple external calls suffered from delay buildup, and even small round trips had a noticeable impact.
In one case, the agent felt unresponsive, not because the model was slow but because of network overhead and cloud cold starts. That forced us to rethink execution placement.
You’ve also explored edge deployments. Where have you seen edge computing outperform expectations, especially in terms of speed, privacy, or cost?
In real-time and persistent scenarios, edge computing has exceeded expectations. Running microVMs closer to users significantly reduced latency and lowered infrastructure costs, especially when agents remained warm across sessions. It also allowed us to keep data local, which improved both compliance and user trust.
You often advocate for a hybrid approach. Based on your experience, what does a smart balance between edge and cloud look like, and what does it take to get it right?
A common mistake is assuming a single cloud deployment is always the best option. Teams often overlook how much latency, cold start time, and tool response impact the agent’s overall performance. While cloud offers powerful centralized capabilities, that doesn’t preclude a more distributed cloud architecture that leverages edge locations. By the time teams realize they need more control over execution, it is already affecting user experience or cost.
Many teams are just starting to build with AI agents. From your own journey, what’s one mistake you see people make when choosing where to run their agent infrastructure?
I am excited by the move from single-agent demos to persistent, multi-agent ecosystems that collaborate and adapt over time. Supporting that shift will require infrastructure that goes beyond stateless APIs. We will need architectures that are goal-oriented, stateful, and distributed, capable of tracking context, sharing memory, and moving seamlessly between edge and cloud environments.
Looking ahead, based on everything you’ve built so far, what excites you most about the future of AI agents, and how do you think infrastructure needs to evolve to support what’s coming next?
What excites me most is the shift from single-agent demos to persistent, multi-agent systems that collaborate over time. To support that, infra needs to evolve from stateless, request-driven architectures to something more goal-oriented, stateful, and distributed. Agents will need to move, share memory, and recover context across cloud and edge seamlessly.
Mathis Joffre is an engineering leader specializing in AI infrastructure. He previously contributed to AI platform development at one of Europe’s largest cloud providers and now focuses on runtime design, model orchestration, and developer tooling for autonomous systems. He is passionate about making cutting-edge research deployable in production at scale.
