# We ran 5,779 AI agent tests on 878 websites. The harness decided what agents found _Five real agent harnesses and a plain model answered the same eight questions about 878 websites. The harness decided how much of each site the agent reached._ **Published:** 2026-11-19 **Source:** https://www.startuphub.ai/ai-news/artificial-intelligence/2026/ai-agent-harnesses-tested-on-878-websites --- Since October 1, StartupHub.ai's [Live Agent Run](/agent-readiness) has sent five real AI agent harnesses to 878 websites: Claude Code, Codex CLI, OpenCode, Cursor and Gemini CLI. Each run asked the same eight questions about a site, 5,779 runs in all. A plain model with a single fetch tool answered the same questions as the control. The result wasn't close. The plain model found 2.49 of the eight answers, while the four harnesses with more than 800 runs each found between 3.69 and 4.25 on the same sites with the same questions, which is 48% to 71% more. ## The harness, not the model, decides how much of a site an agent reads An AI agent harness is the program wrapped around a model. It decides which tools the model can call, how many pages it may fetch and when it stops looking. Swap the harness and the same kind of model reaches a different amount of the same website. The plain model fetched 1.3 pages per run on average. Then it stopped. Codex CLI fetched 5.7 pages for 3.69 answers and Claude Code fetched 5.8 pages for 3.75 answers, while OpenCode got 3.81 answers from 5.0 pages and Cursor reached 4.25 answers from 8.2 pages. Gemini CLI found 5.98 answers, but it ran on only 59 sites, too few to rank it against the others. Answers found per run, out of eight questions. Plain model, one fetch tool: Answers found 2.49. Codex CLI: Answers found 3.69. Claude Code: Answers found 3.75. OpenCode: Answers found 3.81. Cursor: Answers found 4.25. Gemini CLI (59 sites): Answers found 5.98. 5,779 Live Agent Run runs on 878 websites, October 1 to November 17, 2026. StartupHub.ai data. ## More answers took more time More answers, more time. A plain-model run took about 2 seconds at the median, while OpenCode took 20 seconds, Codex CLI 31, Claude Code 43 and Cursor 47, and Gemini CLI took 174. For a person waiting on an agent, that is the difference between an instant reply and one that browses for most of a minute. | Agent | Runs | Websites | Answers (of 8) | Pages fetched | Median time | | --- | --- | --- | --- | --- | --- | | Plain model, one fetch tool | 2,236 | 847 | 2.49 | 1.3 | 2s | | Codex CLI | 885 | 878 | 3.69 | 5.7 | 31s | | Claude Code | 883 | 876 | 3.75 | 5.8 | 43s | | OpenCode | 869 | 862 | 3.81 | 5.0 | 20s | | Cursor | 846 | 839 | 4.25 | 8.2 | 47s | | Gemini CLI | 59 | 59 | 5.98 | 7.5 | 174s | ## Higher Agent Readiness scores, more answers Sites that score higher on [Agent Readiness](/agent-readiness) gave the harnesses more to work with. Harness runs on sites scoring 80 to 100 found 4.75 answers on average, against 4.11 on sites scoring 60 to 79, across 750 and 867 runs. Below 60 there were only 32 harness runs, so those bands show a direction rather than a measurement: 4.00 answers for scores of 40 to 59 and 2.38 below 40. Answers found by agent harnesses, by the site's Agent Readiness score. 0 to 39: Answers found 2.38. 40 to 59: Answers found 4. 60 to 79: Answers found 4.11. 80 to 100: Answers found 4.75. Harness runs on sites with a readiness score: 16, 16, 867 and 750 runs. The two lowest bands are small samples. ## What it means for a website Three things follow for anyone who wants agents to find the right facts about a website. - **A plain-model check understates what agents can reach.** A test that sends a model with one fetch found about 60% of what Cursor found on the same sites. A score built that way describes a weaker reader than the agents people use. - **Put the facts a few links from the homepage.** The harnesses read five to eight pages per run. Pricing, contact details, documentation and what the company does should be reachable in that many steps. - **Fix what the readiness scan flags.** Sites scoring 80 or more gave agents 4.75 answers on average, against 4.11 for sites scoring 60 to 79. [Run the scan](/agent-readiness) to see where a site sits, and Live Agent Run to watch the agents try it. ## How we measured The runs date from October 1 to November 17, 2026, on the current version of the task: eight questions about the site, each run scored from 0 to 8 by how many it answered. The websites are the ones people submitted to the Agent Readiness scanner, so they skew toward small companies rather than famous brands. Each harness ran with the same questions, with a model using one fetch tool as the control. Gemini CLI joined later than the others. It covers 59 sites. Our [study of what AI assistants read on startup profiles](/ai-news/artificial-intelligence/2026/what-ai-assistants-read-startup-profiles) covers the other side: which pages assistants fetch for the people asking them. ## Questions ### What is an AI agent harness? The program around a model that decides its tools, how many pages it may fetch and when it stops; Claude Code, Codex CLI and Cursor are harnesses. ### Which AI agent found the most answers about a website? Cursor, at 4.25 of 8 among the four harnesses with 800+ runs; Gemini CLI scored 5.98 but on only 59 sites. ### Why does a plain model find fewer answers? It fetched 1.3 pages per run and stopped, while the harnesses read five to eight pages and followed links to the facts. ### How can I test my website with real AI agents? Run the free Agent Readiness scan, then Live Agent Run, which sends the same harnesses to your site and shows what they found. --- Original analysis from [startuphub.ai](https://www.startuphub.ai), the #1 AI startup directory. © StartupHub.ai. All rights reserved. You may not republish this article in full without a license. Search engines and AI research tools may crawl and summarize for reference. Bulk reproduction or model training on this content requires a license. See https://www.startuphub.ai/terms.