It is the default search for Perplexity's Computer tool
Super Dario
Introducing Search as Code, our new search architecture for AI agents.
It writes Python that calls our search stack directly, instead of looping through function calls one at a time.
Available in the Perplexity Agent API, and now default in Computer.
https://research.perplexity.ai/articles/rethinking-search-as-code-generation
Many users praised Perplexity's Search as Code for AI agents as a natural and effective way to manage context and build learning loops, while some called the idea unoriginal slop or a messy coordination problem without humans.
How does it impact the average developer that is used to is used to hitting an sdk?
It gives developers an Agentic Search SDK of low-level primitives instead of monolithic endpoints, letting agents generate custom Python pipelines in a sandbox. Average SDK users switching to the Agent API or Computer get better efficiency and results by default with fewer roundtrips. Custom agent builders gain task-specific control and scalability but integrate code execution environments. ¹
We’re moving away from search as a web fetch tool call to search as codegen to be future proof in a world where code execution inside agent harnesses is the way to do almost all of our knowledge work.
Doing this lets you compose multi-step primitives far more naturally and be much more adaptable to changes made to the agent harness, as well as benefit from improvements in coding capabilities that are guaranteed to come from the next generation of frontier models.
Introducing Search as Code, our new search architecture for AI agents.
It writes Python that calls our search stack directly, instead of looping through function calls one at a time.
Available in the Perplexity Agent API, and now default in Computer.
https://research.perplexity.ai/articles/rethinking-search-as-code-generation
“Skillify it” is the way you should write most tasks
Write markdown that makes code
Don’t make elaborate Foxconn factories to call agents
Let agents make their own tools, kaizen style
We’re moving away from search as a web fetch tool call to search as codegen to be future proof in a world where code execution inside agent harnesses is the way to do almost all of our knowledge work.
Doing this lets you compose multi-step primitives far more naturally and be much more adaptable to changes made to the agent harness, as well as benefit from improvements in coding capabilities that are guaranteed to come from the next generation of frontier models.
slowly we're all realizing that tools should be called from code, not from within the llm api
Introducing Search as Code, our new search architecture for AI agents.
It writes Python that calls our search stack directly, instead of looping through function calls one at a time.
Available in the Perplexity Agent API, and now default in Computer.
https://research.perplexity.ai/articles/rethinking-search-as-code-generation
we're going beyond traditional tool calls / MCPs to interact with the search stack. codegen is the most natural way for an LLM to drive search: the research tasks our users need require complex pipelines, customized per task.
so we're exposing the search stack as composable low-level primitives in an SDK. for example, from model-generated Python you can: - build custom ranking pipelines - run large-scale wide/grid search - do map-reduce directly over our search index
the other big win is token efficiency. codegen pipelines things together and collapses many agent steps into a single program, resulting into far fewer round trips.
Introducing Search as Code, our new search architecture for AI agents.
It writes Python that calls our search stack directly, instead of looping through function calls one at a time.
Available in the Perplexity Agent API, and now default in Computer.
https://research.perplexity.ai/articles/rethinking-search-as-code-generation
ptc is the way
Introducing Search as Code, our new search architecture for AI agents.
It writes Python that calls our search stack directly, instead of looping through function calls one at a time.
Available in the Perplexity Agent API, and now default in Computer.
https://research.perplexity.ai/articles/rethinking-search-as-code-generation
@AravSrinivas Skillify confirmed
http://x.com/i/article/2061176923531984896
Good ELI 5 by Grok
Perplexity’s new “Search as Code” lets the AI write a short Python script instead of asking for search results one question at a time.
The script runs inside their system and can do parallel searches, custom filters, deduping, and smart processing all in one go.
Faster, fewer back-and-forth steps, and it gets better as AIs improve at writing code. Basically trading slow conversations for a custom research recipe.
It’s code all the way down
Introducing Search as Code, our new search architecture for AI agents.
It writes Python that calls our search stack directly, instead of looping through function calls one at a time.
Available in the Perplexity Agent API, and now default in Computer.
https://research.perplexity.ai/articles/rethinking-search-as-code-generation
@perplexity_ai @PPLXDevs This is so cool ! We explored a similar idea from the "deep search" side: open-source deep search with dynamic search graphs, open models, and API deployment. Fun to see the industry converging on search becoming programmable. https://github.com/Agent-Field/af-deep-research
@perplexity_ai @MinakoOikawa 😎
☝️
🍷
😝🌏
@perplexity_ai @MinakoOikawa 🫶
👻
😷
🍋👇
Perplexity’s new “Search as Code” lets the AI write a short Python script instead of asking for search results one question at a time.
The script runs inside their system and can do parallel searches, custom filters, deduping, and smart processing all in one go.
Faster, fewer back-and-forth steps, and it gets better as AIs improve at writing code. Basically trading slow conversations for a custom research recipe.
We tested Search as Code on deep research (DSQA, BrowseComp, HLE) and wide research benchmarks (WideSearch, WANDR).
It matches or beats every competing system across all five.
@PPLXDevs The traditional tool-calling approach suffers from high latency, manual control flow, and context pollution.
With Search as Code, the model composes search primitives: fanning out queries asynchronously, deduping, filtering, joining, and ranking before results hits its context.
WANDR is our in-house wide benchmark, built to mirror real professional research workloads.
Search as Code scores 0.386 to the next best system's 0.152, and the benchmark is far from saturated.
We're releasing WANDR in the coming weeks.
Wooow, I implemented something like this in my product about half a year ago.
Except it wasn’t just about search - it applied to tool calling in general.
The reason is intuitive: there is far more data on writing code than on tool calling.
(Although you still need to provide some guardrails, like max_char_num or retries)
More info about Search as Code in the Perplexity Agent API docs:
https://docs.perplexity.ai/docs/agent-api/tools/sandbox
@perplexity_ai @PPLXDevs @grok detailed thread on research and application sources so I can hand off to Claude workflow swarm
@perplexity_ai @PPLXDevs if you want to do this in your agents use DSPy and read the RLM paper or just use the open source ax library ax-agent http://axllm.dev
this is the api-era of agents ending. function calling was fixed schemas. code generation is a programmable runtime. your primitives just became a standard library the agent programs against. every computing platform made this move from fixed apis to scripting. agents are doing it in 18 months instead of 18 years