Perplexity AI's Search as Code cuts token use 85% for AI agents

Perplexity AI has introduced Search as Code (SaC), a new architecture that enables AI models to write their own search pipelines in Python instead of calling fixed APIs. Announced on June 6, SaC addresses a fundamental flaw in how AI agents interact with search infrastructure: the standard loop of writing a query, getting results, and iterating was designed for humans, not autonomous systems performing hundreds of rapid searches. By letting models control ranking, filtering, and deduplication through code, SaC eliminates the context-window bloat caused by rigid backend logic.

In benchmark tests, SaC achieved 100% accuracy on a task of tracking 200 high-severity CVEs while consuming only 42,900 tokens — an 85% reduction from Perplexity's own standard pipeline's 288,700 tokens. Competing systems from OpenAI and Anthropic scored below 25% on the same task. On the WANDR broad-research benchmark, SaC showed a 45% absolute improvement over Perplexity's baseline. The architecture operates in three layers: a reasoning model that decomposes tasks and generates Python code, a compute sandbox with a persistent filesystem for cross-turn state, and an agentic search SDK exposing atomic primitives like retrieve, filter, rerank, and dedupe.

For developers, SaC marks a shift from black-box search APIs to programmable infrastructure. Now rolling out in Perplexity Computer, the approach allows AI agents to design custom search strategies — essentially codifying the intuition of a human engineer. This puts pressure on API providers to expose internals as programmable primitives, and favors frontier models with strong code-generation capabilities. As Perplexity notes, 'Code is a powerful medium for orchestrating preexisting capabilities — it can also serve as a gap-filler for capabilities not present in the search stack.'