Talk
Modern AI agents are increasingly limited by traditional tool-calling workflows. This talk explores a new approach that treats search as executable code, enabling more scalable, efficient, and programmable agent systems. We will discuss the underlying architecture, key engineering trade-offs, and lessons learned from deploying this paradigm in production, highlighting its impact on both performance and efficiency.
The Core Problem: Traditional sequential tool-calling floods context windows and scales poorly. We solved this by shifting to a "Search as Code" paradigm - allowing agents to write and execute their own data retrieval programs in isolated, sandboxed Python runtimes.
Concrete Results Shipped: Aleksandr will break down the sandbox design and persistent file system architecture that achieved an 85.1% token usage reduction and reached 100% accuracy on a complex 200+ CVE task, outperforming the next-best baseline by 2.5× on the WANDR production benchmark.
Key Takeaways: Engineering sandboxed environments, managing runtime state, and building fine-grained programmable primitives using a Python/FastAPI backend layer.