Perplexity launches Search as Code: Agents can directly search through Python code According to Beating monitoring, Perplexity announced the full launch of its new search architecture, Search as Code (SaC), in the Perplexity Agent API and self-developed Computer products. The new architecture abandons the traditional serial loop mode where agents rely on multiple rounds of inference for function calling or MCP tool calling. Instead, agents directly generate Python code and call the modular retrieval primitives provided by the Agenetic Search SDK to parallelize thousands of search and processing operations within a single inference cycle. Traditional tool calls face pain points such as high latency, limited control flow, and context pollution. SaC allows large models to run Python code in a secure computing sandbox, dynamically orchestrating parallel queries, deduplication, filtering, merging, and reordering. To address the challenge of transmitting intermediate states across inference cycles, the team adopted an explicit serialization (serde) scheme for persistent file systems, which is more reliable in long trajectory tasks compared to the REPL pattern. At the same time, the development team introduced exclusive Agent Skills prompt word packages with up to 2000 tokens to guide the efficient orchestration of complex retrieval tasks for the large model. In multiple search benchmark evaluations, the SaC architecture has performed excellently. In the self-developed WANDR benchmark test for measuring professional wide range retrieval, the SaC score reached 0.386, which is 2.5 times higher than the suboptimal Anthropic system (0.152) and 45% higher than the non SaC baseline. In real case tests targeting 200 high-risk CVE vulnerability queries, SaC achieved 100% accuracy and reduced token consumption by 85.1% compared to traditional baselines. In the DSQA evaluation, SaC achieved a high score of 0.871 while reducing the cost of each task at a moderate level of reasoning to below $1. [Original link]