Perplexity has launched a new search architecture called Search as Code (SaC), designed to enhance the efficiency and capability of AI agents by allowing them to write Python code that directly interfaces with the search stack, instead of relying on traditional, slower function calls. This new system, now available through the Perplexity Agent API, not only improves response times but also increases cost-efficiency, as demonstrated by benchmark tests where SaC outperformed competing systems on various metrics, including its own scores across the newly developed WANDR benchmarking suite, which measures the ability to handle complex, knowledge-intensive tasks. With the integration of models for decision-making, secure sandboxes for code execution, and a modular Agentic Search SDK, SaC signifies a significant advancement in AI search capabilities.

WANDR: WANDR is Perplexity’s in-house benchmark designed to evaluate wide research workloads that mirror professional tasks. It stresses complex orchestration of search and reasoning across multiple operations. The benchmark is scheduled for public release in the coming weeks.
Perplexity: Perplexity is an AI company that develops search and answer engine products for users and developers. It has launched the Search as Code architecture to enable more efficient agentic workflows. The company makes this available through its Agent API and Computer product for advanced research tasks.
Search as Code: Search as Code is Perplexity’s new reference search architecture that allows AI models to generate and execute Python code for orchestrating search primitives directly. It addresses limitations of traditional function-calling approaches by enabling asynchronous, composable retrieval strategies. The feature is integrated into the Agent API and set as default in Computer.
Perplexity Agent API: Perplexity Agent API is the developer platform that provides access to Perplexity’s agent capabilities and search features. It now includes Search as Code as a core offering for building programmable search pipelines. The API supports tools like sandbox execution alongside the new architecture.

`json
{
“Architecture”: “The architecture combines models as the control plane, secure sandboxes for code execution, and an Agentic Search SDK for direct access to search building blocks.”,
“Benchmarking”: “WANDR focuses on difficult wide research tasks that require careful orchestration of search, compute, and model reasoning.”,
“Product Feature”: “Search as Code allows models to assemble Perplexity’s search stack components via code execution within sandboxes.”
}
`