World's Fair 2025
Building an Agentic Platform — Ben Kus, CTO Box
Overview
This talk explores Box's journey in building an agentic platform, focusing on how agentic approaches can solve complex data extraction challenges that traditional methods and basic LLM calls struggle with. The core thesis is that an agentic abstraction layer provides a clean, flexible, and evolvable architecture for tackling sophisticated AI tasks, moving beyond simple chatbot interactions to orchestrate complex workflows.
Who should watch
- AI Engineers
- Product Managers
- Builders working with unstructured data
- Teams facing challenges with accurate data extraction from complex documents
- Those interested in agentic AI architectures and their applications
Key takeaways
- Generative AI initially improved data extraction from unstructured documents, but struggled with complexity, scale, and accuracy on more intricate tasks.
- Agentic approaches, involving AI agents with instructions, objectives, tools, and memory, offer a robust solution for complex problems like extracting hundreds of fields from lengthy documents.
- An agentic architecture allows for intelligent orchestration, breaking down tasks, using tools for verification, and employing multi-model approaches or LLMs as judges for feedback and refinement.
- This agentic framework simplifies evolution; changes can be made by adjusting prompts or adding verification steps rather than redesigning the entire system.
- Separating the agentic framework from distributed systems design is crucial for managing both intelligent workflows and large-scale data processing.
- Adopting an AI-first, agentic mindset is essential for teams to effectively build and leverage these capabilities.
- Box offers an agent API, allowing programmatic interaction with their agents for various tasks.
- Evaluation involves a mix of standard and challenge-based eval sets, LLM-as-a-judge, and user feedback, rather than solely relying on fine-tuning.
Notable quotes
*The agentic abstraction layer from an architecture perspective is actually quite clean.*
*If you're confronted with a challenge, and it's plausible that a set of AI models could help you solve that problem, then you should build this AI agentic architecture early.*
Unofficial community note. Prefer the recording for nuance.