Search as Code

A search setup for developers building agents on Perplexity: the agent writes its own Python to run many searches at once instead of one at a time. Now the default in Perplexity Computer.

Details

The agent writes Python that calls the Perplexity search stack directly, running many search operations in a single turn instead of one function call at a time.
The Perplexity search stack is broken into small composable pieces (retrieval, ranking, filtering) the model can batch and combine with its own logic, instead of one fixed search service.
The generated code runs in a secure sandbox that keeps files between turns, so an agent can build on its earlier work.
Runs on GPT-5.5 in high-reasoning mode.