The “Mixture of Experts” year-end episode brings together several AI experts to reflect on the major trends of 2025 and make predictions for 2026. The discussion opens with a review of last year’s predictions, particularly around AI agents. While there was an expectation that 2025 would be the “year of the agents,” the reality was more nuanced. Agents have become more capable, moving from specialized, single-purpose tools to more general “super agents” that can reason, plan, and orchestrate multiple tasks using various tools. However, their integration into daily workflows has been so seamless that users often don’t recognize them as agents anymore—they simply see them as AI.

A significant theme is the competition among major AI providers—OpenAI, Google, Anthropic, and others—to become the primary “front door” for these super agents. The experts note that the market is shifting away from a proliferation of specialized agents (akin to traditional apps) toward orchestrator agents that can handle a wide range of tasks. This shift is driving a battle for dominance in browsers, mobile platforms, and other user interfaces, as companies embed their AI agents directly into these environments. The panel predicts that the next year will see further development of agent control dashboards and multi-agent orchestration, making it easier for users to manage and deploy agents across different contexts.

Open source AI also emerges as a major story of 2025. The gap between proprietary and open source models has narrowed significantly, with some open source models now rivaling or surpassing their closed-source counterparts. However, open source still faces challenges in packaging and user experience. While the components are available and often superior, there is not yet a standard, user-friendly way to assemble them into cohesive, high-quality applications. The experts are optimistic that 2026 will see the emergence of more refined developer patterns, middleware, and orchestration frameworks, making open source AI more accessible and competitive across a broader range of domains.

AI hardware is another focal point, with 2025 marked by persistent supply constraints and a growing divide between large-scale, high-performance clusters and efficient, edge-capable devices. Companies have had to optimize their strategies around compute availability rather than just model capability. Looking ahead, the panel expects 2026 to be defined by the tension between “frontier” models (large, powerful, but resource-intensive) and efficient models (smaller, hardware-aware, and able to run on modest devices). There is also anticipation of new hardware innovations, including specialized chips for agentic workloads and increased competition to Nvidia’s dominance.

Finally, the episode explores the evolution of multimodal AI—models that can process and integrate text, images, audio, and more. Multimodal capabilities are becoming standard, but the focus is shifting toward modularity and orchestration, allowing different specialized models to work together within a workflow. Human-in-the-loop approaches and adapter-based architectures are highlighted as ways to improve accuracy and flexibility, especially for complex or enterprise-specific tasks. The experts predict that 2026 will see further advances in multimodal AI, with more sophisticated orchestration layers and a clearer understanding of how to best combine and deploy these capabilities in real-world applications.