Two days, built for engineers.

Search as Code: The Infrastructure Behind Long-Running Agents

Aleksandr Nikolenko

Once search becomes something an AI agent can program rather than just call, the underlying infrastructure requirements change entirely. This session demonstrates how moving from fixed, serial tool calls to long-running, sandboxed pipelines enables agents to orchestrate thousands of retrieval operations. We will examine how this specific architecture achieved 100% accuracy on a complex 200+ CVE task using 85.1% fewer tokens than serial baselines, and outperformed the next-best system by 2.5× on the WANDR production benchmark. Using Perplexity’s Search as Code as a technical case study, this talk breaks down the exact infrastructure that makes these results possible. We will unpack the core components required for agents to build custom retrieval pipelines: secure sandboxes for executing code, persistent state and filesystems to avoid polluting model context, and programmable search primitives. The session will detail why engineering decisions around sandbox design, state management, and SDK granularity are just as critical to an agent's success as the underlying LLM itself.