The WANDR Benchmark and the Myth of the 10x Research Agent

What is the WANDR benchmark and why does it matter for AI research?

The WANDR (Wide ANd Deep Research) benchmark is an open-source evaluation framework released by Perplexity in July 2026. It measures the ability of AI agents to perform large-scale, evidence-backed data collection tasks, requiring them to search broadly for entities and investigate deeply to provide verifiable proof for every claim.

The promise of autonomous research agents is currently facing a cold reality. While marketing narratives suggest AI can replace analysts in due diligence and market research, the data tells a different story. Perplexity's WANDR benchmark release reveals a sobering gap: even the most advanced systems fail 87% of complex tasks requiring rigorous evidence verification at scale.

Key Takeaways

• Reliability Crumbles at Scale: As research tasks grow in volume, agent accuracy drops drastically due to context rot.
• The Discovery vs. Proof Gap: Finding a link is easy; extracting a specific excerpt that actually proves a claim is where most systems fail.
• Low F1 Scores: The leading system (Search as Code) achieved only a 0.133 Hard F1 score on the full benchmark suite.
• The Cost of Depth: High-quality research is resource-intensive, with some tasks costing hundreds of dollars without guaranteeing accuracy.

The 'Wide and Deep' Problem: Why Agents Fail at Due Diligence

Most existing benchmarks ask if a model can answer a single question correctly. WANDR asks something much harder: Can the agent find all 70 companies meeting a specific criterion, verify a specific fact about each, and cite a page that actually proves it?

This introduces the concept of a qualification key hierarchy. A typical task requires the agent to navigate a decision tree: Company -> Employee -> Evidence URL. If one branch breaks, the entire record is void. According to AlphaSignal's analysis, this structure punishes agents for inconsistency, leading to scores far lower than those seen in standard chatbot evaluations.

Performance Comparison: Leading Research Agents (July 2026)

The data shows that even the winner, Perplexity, is far from solving the problem. A score of 0.133 means the system earned full credit for only about one in seven required records.

The Scale Tax: How Wide Discovery Leads to Deep Hallucination

One of the most fascinating findings in the AI Insiders report is that scale acts as a poison for AI agents. In small tasks, accuracy is reasonable. But as soon as an agent is required to collect hundreds of records, 'context rot' sets in. The agent begins to cut corners on evidence verification to meet volume demands.

The problem isn't finding the right page. Data shows only a small percentage of pages are unusable. The real failure (between 57% and 86%) is in the excerpt extraction. The agent submits a relevant link, but the information within that link doesn't actually prove what the agent claims it proves.

Operationalizing the Gap: Building Systems That Handle Agentic Failure

For organizations, the takeaway isn't to stop using AI, but to change how it's deployed. Instead of relying on an agent to submit a final report, build 'human-in-the-loop' systems.

1. Define Stop Criteria: Don't let an agent run indefinitely. Set checkpoints where a human verifies the quality of the first few evidence chains.
2. Cross-Verification: Use one agent for research and a second, different model to audit the citations provided.
3. Task Decomposition: Instead of asking for a 'full competitor map,' break the task into small sub-tasks where the agent is less likely to lose focus.

The future of AI research isn't a magic agent that does everything, but an infrastructure where every piece of information is immediately verifiable by the user. As the explainx.ai blog notes, this benchmark is a wake-up call for anyone who thought human analysts could be replaced with a single click.

Sources

• WANDR Benchmark: Evaluating Research Agents That ... — web
• WANDR Benchmark Open Source — July 2026 | explainx.ai Blog | explainx.ai — web
• Perplexity's Open-Source WANDR Benchmark Exposes How Badly Research Agents Fail at Scale | AlphaSignal — web
• Understanding Perplexity WANDR Benchmark Methodology — web
• Perplexity's WANDR Benchmark Shows Research Agents Still Fail at Scale — AI Insiders — web

What is the difference between Soft F1 and Hard F1 in AI research?

Soft F1 grants partial credit for partial success, such as finding the right company but failing to provide a perfect evidence excerpt. Hard F1 requires the entire data tree to be perfect and backed by evidence to receive any points.

Can AI agents perform due diligence independently?

According to the WANDR benchmark, the answer is currently no. Agents fail the majority of complex tasks and require heavy human oversight to ensure cited sources actually support the factual claims made.

Why do research agents become less reliable as the task size increases?

This is known as 'context rot.' As an agent processes more data and performs more search actions, the likelihood of cumulative errors, duplicates, and losing the ability to verify every detail against the original source increases.

3 things to remember

• Current AI research excels at finding leads but fails at systematic verification at scale.
• The gap between 'looks right' and 'proven by source' is the biggest hurdle for research agents in 2026.
• Never rely on AI summaries without direct access to original excerpts and manual spot-checking of the evidence trail.