Search as Code: The Programmable Future of Agentic Retrieval

Perplexity's Search as Code (SaC) architecture replaces rigid search APIs with a programmable SDK that AI agents can orchestrate through code generation. This hybrid approach combines LLM reasoning with deterministic compute sandboxes to execute complex, parallelized retrieval strategies tailored to specific tasks. Benchmarks demonstrate that SaC dramatically improves performance on difficult research tasks while significantly lowering token costs and context overhead.

Key Points

• Traditional monolithic search APIs are insufficient for complex agent tasks that require thousands of retrieval operations and fine-grained control.
• Search as Code (SaC) deconstructs search into an Agentic Search SDK, allowing models to orchestrate retrieval pipelines via generated Python code.
• The architecture utilizes secure compute sandboxes to execute deterministic operations like batching, filtering, and joining, which reduces context bloat.
• SaC achieves state-of-the-art results on 'wide research' benchmarks, outperforming competitors like OpenAI and Anthropic by significant margins.
• The system establishes a new cost-performance frontier, delivering higher accuracy at lower costs compared to traditional search-integrated LLM calls.

Sentiment

The community reaction is mixed but generally constructive. Commenters are interested in the idea's potential for agentic code search and context reduction, yet they remain skeptical about whether the architecture is novel, whether it meaningfully improves on rich APIs or existing agent workflows, and how it handles operational limits. Overall, HN neither fully endorses nor dismisses the article; it treats the proposal as plausible engineering with unresolved product and systems questions.

In Agreement

• Agents could benefit from turning repeated search-tool loops into a single sandboxed program that fans out retrieval, filters results, deduplicates hits, and returns only relevant context.
• The approach maps well to grep-based code search because an initial generated program could build a repository map before more targeted searches, reducing context pollution without requiring prior codebase knowledge.
• Specialized retrieval primitives may be more appropriate than forcing free-form document search into relational query models, especially when search is a core competency rather than an add-on.
• Moving intermediate retrieval work into a sandbox fits the broader pattern of bringing computation closer to the data and can make agent workflows more efficient when search work is substantial.

Opposed

• Several commenters see the proposal as a rebranded query language or an old database pattern with AI-specific framing rather than a fundamentally new search paradigm.
• Some argue that a secure SQL interface or a sufficiently capable search API might provide the same control with less complexity.
• For coding agents, skeptics note that existing iterative search, parallel tool calls, and subagent patterns may already solve much of the exploration problem, making generated search scripts less valuable outside slow or compute-heavy searches.
• Generated Python retrieval plans could create difficult support and limits problems because inefficient code may be produced by the model while customers are not expected to understand or fix it.