Perplexity AI Releases WANDR: An Open Benchmark for Research Agents That Need to Search Wider and Deeper

Research agents are already handling the actual information work today. Teams submit competitive maps, due diligence, and literature reviews to them. However, most benchmarks evaluate a single response, not large evidence-based collections. Confusion is directed at the gap with a new open benchmark.

Confusion is released WANDR (Wide AND Deep Research). It is an open stop sign and inspection harness. It is designed with nearly 500 practical, challenging data collection tasks for information work. WANDR is the extended child of Perplexity’s DRACO for intensive research. DRACO asks if the agent produces an accurate, complete, objective long-form report. WANDR instead asks if it can build a larger collection with evidence.

What is WANDR

At its core, WANDR explores two quests together. Broad means finding a large, often open set of suitable businesses. Deep means investigating every organization sufficiently to support each claim with evidence. Combining the two changes the problem to agents. A few compelling examples are not enough here. A polished narrative built on imperfect research also falls short.

To capture this, WANDR uses composables List of qualification key categories. One job may ask company(n) -> employee(m) -> url(k). This means n eligible companies, m employees each, and k supporting pages for each. Every complete path through the tree is independently verified. The same structure can represent a flat list, a nested search, or a matrix.

Example of Concrete Work

To support that list, consider the exemption ceo_cfo_appointments work. It solicits at least 70 US-based companies. Each must have a CEO or CFO appointment first announced between March 1 and April 30, 2026. For each, the agent provides one approved appointment page. A subtask adds a listing authority page for each company. Altogether, the work requires 140 source-based records.

Specifically, these two sections and one submitted record look like this:

# Task hierarchies
company(70) -> company_appointee(1) -> url(1)   # 70 appointment records
company(70) -> url(1)                           # 70 listing records

# One record the grader re-fetches and re-checks (values are illustrative)
{
  "item":     "Example Corp - new CFO",
  "url":      "
  "excerpts": ["Example Corp today named Jane Doe as Chief Financial Officer, effective April 2026."],
  "answer":   "Jane Doe appointed CFO; announced April 2026"
}

Realistic Works, Produced to Scale

Beyond single examples, WANDR builds its operations on real-world applications. It starts from undetected patterns observed in production, not from synthetic instructions. A semi-automated pipeline then turns those patterns into jobs. The pipeline has four stages: seeding, approval, entry, and maintenance. It uses a branched author-critic loop with a synthesis mechanism.

As a result, the average job requests 50 members and 245 records in total. Of all 500 jobs, WANDR calls for 170,495 source-based records. The tasks are divided into 167 low, 166 medium, and 167 high difficulties. The difficulty depends on the function of each record, not on the scale alone.

How WANDR Learns Agents

Unlike fixed answer keys, WANDR weighs each claim against stated evidence. All records contain an item, a URL, selected quotes, and a response. The compiler reloads the page during the test. Checks if the page is usable and wide. Then it verifies the quotes that actually appear and support all the requirements.

These binary record decisions then proceed to the classification stage. Accuracy measures the quality of what is delivered by the system. Remember it measures the quality-corrected finish, filling any deficiency with zero. It’s soft points that give partial credit to incomplete members. It’s difficult scores are calculated only for members whose full bottom line is correct.

Benchmark results

Using that approach, Perplexity ran six production systems for a total of 500 jobs. Its Search as Code (SaC) program is leading. However, neither system comes close to solving the benchmark.

With more effort, Perplexity reaches a soft 0.447 F1 xhigh preparation. The cost in all settings exceeds by more than four orders of magnitude. From $0.03 per transaction to $324.83 per transaction.

Beyond the leaderboard, four findings stand out. First, partial progress is normal, but complete coverage is not. Every system shows soft recall under soft precision. Second, scale compounds the problem greatly. Deeper sections are more damaging, as each branch adds a point of failure. Third, availability is the first bottleneck of the structure. High-level detection completions ranged from 0.611 to 0.951 for all systems. Under-delivery, not double consolidation, explains the missing volume. Fourth, finding a usable page is usually easy. Turning it into a complete proof is the hard part. Disturbingly, 41.4% of pages miss the main requirement. Also, 57.5% of the citations fail to support the full claim. Its soft F1 drops from 0.531 under the retrieval check only to 0.363 under the full resolution.

Notably, Search as Code is well-suited to this type of work. The agent can perform retrieval, sorting, extracting, joining, repeating, and configuring logic as a program. Deterministic compute then handles iterative operations outside the context of the model.