Perplexity Adds Wide Research to Its Agent API and Releases the 500-Task WANDR Benchmark
Perplexity has added a Wide Research preset to its Agent API and released WANDR, a 500-task benchmark for evidence-backed research work.
Perplexity has made a Wide Research preset available in its Agent API and released WANDR, a 500-task benchmark designed to test whether AI systems can complete broad, evidence-backed research assignments rather than merely locate a few relevant documents.
The product change is aimed at a class of work where the central requirement is coverage: finding every result that meets stated criteria, enriching those results with additional facts, distinguishing similarly named entities, and supplying evidence for each conclusion. In announcing the release, Perplexity CEO Aravind Srinivas described ordinary “narrow search” as largely solved and framed the harder problem as wide research conducted at scale.
The distinction matters because a useful answer to a research request is not always a good research deliverable. A model may identify several plausible examples from a large universe of candidates, but still miss qualifying entities, merge records that refer to different people or companies, or provide citations that support only part of its final list. Those failures are especially costly in workflows such as market mapping, vendor screening, public-record research, competitive analysis, and literature or policy review.
What Perplexity is adding to the Agent API
Perplexity says the Wide Research preset uses its “Search as Code” architecture. The stated design is to let a model create a research plan once and then execute that plan deterministically across a larger set of searches and retrieval operations, rather than repeatedly carrying every intermediate result through the model’s context window.
That is a different technical emphasis from a conventional chat-based research flow. In a narrow task, an agent can often search, read a small number of pages, and synthesize an answer within one evolving conversation. A wide task can require the same sequence of checks for dozens or hundreds of candidates: identify candidates, test the eligibility rule, extract fields, fetch supporting pages, resolve duplicates, and produce a structured output.
Perplexity’s claim is not that a model no longer needs to reason about the assignment. The model still needs to determine what to search for, what constitutes a qualifying result, which evidence is sufficient, and how the final output should be organized. The change is in how that plan is executed: repeated operations can be handled as a programmatic research workflow instead of forcing a growing list of raw results back into the model’s working context.
In practical terms, that approach is intended to reduce two common pressures in broad research jobs. First, context can become crowded with records that have already been processed but must remain available for comparison. Second, a long agent run can drift when it has to restate its criteria or re-evaluate the same kind of record many times. A deterministic execution layer cannot by itself guarantee complete coverage, but it can make the repeated portion of a research plan more consistent and more inspectable.
The announcement does not establish a universal definition of “every” qualifying result. Completeness always depends on the available sources, the exact inclusion criteria, and whether relevant material is publicly retrievable. The more concrete claim is that the API now exposes a preset intended for assignments where coverage and per-record evidence are first-class requirements.
WANDR turns broad research into a testable task
WANDR is Perplexity’s accompanying benchmark for “wide and deep research.” Its public repository describes the benchmark as structured, high-volume information work requiring broad discovery, extensive enrichment, systematic extraction, precise entity disambiguation, and evidence-backed answer synthesis.
Those requirements make WANDR notably different from benchmarks that primarily score a single answer or a short chain of reasoning. A system must handle a set of entities and produce the required files or structured outputs for a task. The benchmark’s source tasks include task trees, schemas, prompts, scoring configurations, and task-owned artifacts. The source definitions are then rendered into self-contained task packages.
Each generated WANDR package includes the solver instructions, task metadata, a public Docker environment, evaluator material, and an ordered task manifest. The repository says the task-local evaluator is deliberate: a task can be published, downloaded, and verified without depending on a separate WANDR package at runtime. That makes the benchmark more portable, while also making the evaluation contract more explicit.
The evaluation process is designed to inspect more than whether a final answer sounds convincing. After an agent produces its required output files, WANDR’s verifier can fetch submitted pages, normalize entities, deduplicate records, and judge the result. The repository distinguishes a completed run that receives a zero reward from a verifier error: a zero is a scored outcome, while an error means no valid score was produced.
This is important for research-agent evaluation because an apparently polished narrative can obscure operational mistakes. A system can write fluent prose while failing to return a required record, citing a page that does not support a field, or duplicating the same organization under multiple names. A verifier that checks task-specific files and scoring material is better suited to detect those failures than a broad preference judgment alone.
A benchmark built around reproducibility, not just prompts
Perplexity has published WANDR under the Apache-2.0 license. The repository separates the editable task sources from the generated benchmark packages and from the generic agent adapter used to run remote endpoints. That separation gives researchers and developers clearer boundaries for changing task semantics, translating tasks into the Harbor benchmark format, and testing agents against the resulting packages.
The project also includes configurations for a one-task smoke test, an all-provider smoke test, a two-task validation run, and a full benchmark run. Its documented provider matrix includes OpenAI, Anthropic, Perplexity, Exa, Parallel, and Gemini integrations. The repository cautions that full evaluations make paid solver, fetch, and judge API calls and that the benchmark harness does not impose a spending cap.
That warning is a useful reminder of what wide-research evaluation measures. The benchmark does not reduce a complex research job to a cheap one-shot prompt. Evaluating broad discovery, page retrieval, evidence validation, and structured outputs can require substantial model and web-access usage, especially when the same task is run across multiple systems.
The project’s Relay component records lifecycle events, tool activity, usage, produced files, prompts, final messages, and normalized result metadata for each trial. It also requires declared output files to be present and non-empty before a verifier starts. Those mechanics make it possible to inspect whether a result failed because an agent did not produce the requested artifact, because the output was malformed, or because the submitted research did not satisfy the task’s scoring rules.
The practical change for developers
For developers building research agents, the Wide Research preset is most relevant when the output is a dataset, a reviewed candidate list, or an evidence trail—not simply an answer paragraph with a few links.
A procurement workflow, for example, may need every supplier matching a region, certification, product category, and size threshold, with a source supporting each field. A policy researcher may need all relevant filings in a defined time period, with duplicate entities resolved and each item classified against a fixed rubric. In both cases, finding a few good examples is insufficient; the workflow needs repeatable operations and a result that can be checked record by record.
WANDR provides a public way to test that distinction. Its task structure, verifier design, and published artifacts give developers something more concrete than an anecdotal demonstration when comparing research-agent systems. Perplexity’s API preset, meanwhile, presents Search as Code as an execution model for handling the repeated work such tasks generate.
The release does not prove that any single model has solved comprehensive research. It does make the underlying standard more explicit: broad research systems should be judged on coverage, extraction quality, entity resolution, output contracts, and evidence—not only on whether their final prose is persuasive.
Sources
Share