Perplexity launches Search as Code for faster agentic retrieval

Perplexity is pushing its search stack one level deeper into the infrastructure layer. Search as Code lets agents write Python that calls the search system directly instead of stepping through sequential tool calls, a change aimed at cutting latency and context overhead when a single Perplexity Computer task can trigger hundreds or even thousands of retrieval operations in minutes.

That is more than a product tweak. It turns search into something closer to an analyst’s workflow, where an agent can define a task-specific retrieval strategy inside its own harness, then query, filter, and iterate without waiting for a human-style loop to unwind one request at a time. The upside is speed and tighter grounding. The downside is also built in: Perplexity says code-generation-based search can add its own latency and context cost, especially when the system has to generate serialization code, which is why it says well-designed utility functions matter.

The launch also fits into a rapid expansion of Perplexity’s search stack. The Search API arrived on September 25, 2025, with Perplexity saying it powered a continuously refreshed index covering hundreds of billions of webpages. In its research on that system, Perplexity said the architecture leaned on hybrid retrieval, distributed indexing, multi-stage ranking, and dynamic parsing, and that it handled 200 million daily queries. The company later launched the Agent API on March 11, 2026 as a managed runtime for agentic workflows, replacing a model router, search layer, embeddings provider, sandbox service, and monitoring stack with a single integration point. It is model-agnostic, supports model fallback chains, and exposes built-in tools including web_search and fetch_url.

Perplexity Computer, which launched on February 25, 2026, supplies the runtime where that architecture becomes visible in practice. The company says Computer can reason, delegate, search, build, remember, code, and deliver, while running asynchronously, creating sub-agents, and operating in an isolated compute environment with a real filesystem, browser, and tool integrations. Perplexity’s broader research points in the same direction: Search as Code outperformed prior methods on benchmarks such as DSQA and WideSearch, while Perplexity Deep Research posted state-of-the-art results on external benchmarks including Google DeepMind’s DeepSearchQA and Scale AI’s ResearchRubrics.

Taken together, the launches show Perplexity treating search less like a consumer feature and more like the core runtime for agentic systems. If agents are now writing retrieval logic in code, the competition is no longer just about who answers fastest. It is about which sources are easiest for machines to surface, trust, and reuse under production pressure.