Agentic Web Navigation Benchmark

We asked each agent to complete five realistic web tasks on live sites:

1. Find a specific product on an e-commerce site and add it to cart.
2. Book a meeting slot on a scheduling page.
3. Locate a buried policy document and extract a clause.
4. Submit a support ticket with the correct category.
5. Recover from a 404 by finding the moved page.

Agents were scored on task completion, correctness, and whether they handled errors gracefully.

Criterion	Weight	Max points
Task completion	50%	50
Correct data extracted/entered	25%	25
Error recovery	15%	15
Efficiency	10%	10

• Live sites, not mock pages.
• Agents used browser tools or built-in web access.
• Each task was attempted up to three times; best run scored.
• We did not modify sites; real-world variability is part of the test.

• Manus completed the most tasks end to end, including recovery from UI changes.
• Devin planned carefully but was slower; excelled at structured workflows.
• Claude Code worked best with explicit step-by-step direction.
• OpenClaw needs closer browser-tool integration for open-ended navigation.

• Live sites change; results may not replicate exactly.
• Some agents rely on specific browser providers, which affects performance.
• Ethical and terms-of-service considerations limit the sites we can test.

• Manus: delegating complex web tasks end to end.
• Devin: structured, multi-step web workflows that need planning.
• Claude Code: web tasks embedded in a coding or terminal workflow.
• OpenClaw: self-hosted agents with custom browser tooling.