Lin Qiao, CEO and co-founder of Fireworks AI, opened her talk by emphasizing that 2025 is the year of AI agents. She highlighted the diverse range of agents Fireworks has onboarded, including coding agents that transform software development, document agents for collaborative work, sales and marketing agents that target ideal customer profiles, hiring agents that streamline candidate selection, and personalized customer agents that enhance customer service. These agents span various industries such as insurance, healthcare, and education, demonstrating the broad applicability and growing adoption of AI agents.
Lin noted that successful companies building AI agents share two key principles. First, they do not treat AI models as mere commodities or utilities but integrate them deeply into their product design. Second, they implement a product-model co-design process, creating a “data flying wheel” where data collected from product usage continuously improves the model, which in turn powers better products. This cyclical process helps companies develop unique intellectual property and gain competitive advantages by tightly aligning models with their specific product needs.
The talk then addressed the complexity of building AI agents, which involves three main stages: build, customize, and scale. During the build stage, developers face a rapidly evolving landscape of open models with varying capabilities and performance characteristics. Fireworks simplifies this by providing access to over 200 optimized models through a secure, scalable API platform, allowing developers to focus on building applications without constantly chasing new model releases. This approach helps manage the excitement and exhaustion associated with the fast pace of AI model development.
Customization, the second stage, involves fine-tuning models through reinforcement learning to better align with evolving product requirements. Lin explained that reinforcement learning mimics human learning by using rewards and feedback to improve performance over time. She highlighted the challenges of setting up environments for reinforcement learning, especially for complex multi-turn interactions, and described how Fireworks supports this process with managed training runs, observability tools, and integration with existing business logic and APIs. This enables developers to continuously improve their AI agents using real-world data and feedback.
Finally, Lin discussed scaling AI agents to millions or billions of users, which requires sophisticated infrastructure management. Fireworks operates across multiple cloud providers and regions, leveraging a proprietary 3D optimizer that balances model quality, speed, and cost by exploring over 100,000 configuration options. This tailored approach ensures optimal performance for specific applications. She also mentioned Fireworks’ collaboration with MD hardware to enhance model performance, underscoring the company’s commitment to pushing the boundaries of AI agent technology. Overall, Lin’s talk provided a comprehensive overview of the challenges and solutions in building, customizing, and scaling AI agents with Fireworks AI.
