--- title: Computer-use models will agentify the web, not APIs —... description: A talk from Dhruv Batra of Yutori argues against the popular idea that AI agents will interact with the web primarily through APIs (MCP servers, payment... canonical: https://daily.dev/posts/computer-use-models-will-agentify-the-web-not-apis-dhruv-batra-yutori-fnoivacum twitter:card: summary_large_image twitter:site: @dailydotdev og:type: website og:site_name: daily.dev og:title: Computer-use models will agentify the web, not APIs — Dhruv Batra, Yutori | daily.dev og:description: A talk from Dhruv Batra of Yutori argues against the popular idea that AI agents will interact with the web primarily through APIs (MCP servers, payment... og:url: https://daily.dev/posts/computer-use-models-will-agentify-the-web-not-apis-dhruv-batra-yutori-fnoivacum og:image: https://api.daily.dev/og/posts/FNOIVaCum.png og:image:alt: Computer-use models will agentify the web, not APIs — Dhruv Batra, Yutori og:image:width: 1200 og:image:height: 630 og:locale: en --- > ## Documentation Index > Fetch the complete documentation index at: https://daily.dev/llms.txt > Use this file to discover all available pages before exploring further. # Computer-use models will agentify the web, not APIs — Dhruv Batra, Yutori **[AI Engineer](https://daily.dev/sources/aidotengineer)** · 21 min read · 0 upvotes · 0 comments ## Summary A talk from Dhruv Batra of Yutori argues against the popular idea that AI agents will interact with the web primarily through APIs (MCP servers, payment protocols, etc.). He contends that only the head of the web (major sites) will offer APIs, while the long tail — restaurant menus in scanned PDFs, school district procurement pages requiring FOIA requests — will never expose structured endpoints because institutions change infrastructure too slowly. Instead, computer-use models that see pixels and click buttons like humans will be the way agents access the long tail web, since the web was built for human eyes and much of its content (like stock status or scores) is rendered dynamically rather than present as static HTML text. He presents Yutori's Navigator model, including a version N-1.5 hitting 97% on the Online Mind2Web benchmark, and highlights improvements in latency and cost (80 cents vs $230 per task) compared to frontier models like Opus and GPT. His conclusion: agentic access to the long tail web will emerge from ever-cheaper, ever-faster computer-use agents piling another layer of automation on top of the existing messy web, rather than from a wave of new APIs. ## Full article daily.dev links to this article rather than hosting it. Read it at the original source: ## Questions this post answers ### Why can't AI agents just use APIs or MCP servers to access most websites? Most of the web, especially the long tail beyond popular sites, will never expose APIs because the underlying institutions change infrastructure very slowly. Examples include restaurant menus stored as scanned PDFs or JPEG galleries, and school districts requiring Freedom of Information Act requests just to retrieve procurement documents by email and Google Drive. Only the most popular head-of-distribution websites are likely to offer real APIs. _Developers weighing agent architecture choices can follow this API-versus-computer-use debate on daily.dev._ ### Why can't a coding agent just read a website's HTML to get information instead of using vision? Reading raw HTML often fails because much of what's displayed on a page, like stock status or a live sports score, is not present as static text but is rendered dynamically from a JSON response fetched asynchronously after page load. For example, an in-stock/out-of-stock indicator may only exist as a numeric quantity variable with a separate script controlling grayed-out styling, with no literal 'sold out' text anywhere in the HTML. _Teams debugging why scrapers miss dynamic content can track these browser-rendering gotchas on daily.dev._ ### How much does it cost to run a computer-use agent task compared to large frontier models like GPT-5.5 or Opus 4.7? A smaller, specialized computer-use model called Navigator N-1.5 costs about 80 cents per task for roughly 20-30 interaction steps, compared to about $230 for the same task using trillion-parameter-class frontier models. Navigator N-1.5 also scored 97% human-eval accuracy on the Online Mind2Web benchmark, with only 8 of 300 trajectories judged incorrect, while offering lower per-step latency than larger models. _Anyone comparing agent runtime costs across models can keep tabs on these benchmarks via daily.dev._ ## Similar posts on daily.dev - [What Building AI Agents Inside a Production Startup Taught Me About the Future of the Web](https://daily.dev/posts/what-building-ai-agents-inside-a-production-startup-taught-me-about-the-future-of-the-web-vvgbb49hm) · SD Times · 1 upvotes · 0 comments - [Headless AI will make every product a platform. Here’s what to consider when designing for this.](https://daily.dev/posts/headless-ai-will-make-every-product-a-platform-here-s-what-to-consider-when-designing-for-this--v9ypwkblk) · Medium · 1 upvotes · 0 comments - [How to stop holding AI agents back](https://daily.dev/posts/how-to-stop-holding-ai-agents-back-86grinkth) · The Next Web · 0 upvotes · 0 comments --- Tags: [#automation](https://daily.dev/tags/automation), [#ai-agents](https://daily.dev/tags/ai-agents), [#mcp](https://daily.dev/tags/mcp), [#crawling](https://daily.dev/tags/crawling) [View this post on daily.dev](https://daily.dev/posts/computer-use-models-will-agentify-the-web-not-apis-dhruv-batra-yutori-fnoivacum) ```json {"@context":"https://schema.org","@graph":[{"@type":"Organization","@id":"https://daily.dev/#organization","name":"daily.dev","url":"https://daily.dev","logo":{"@type":"ImageObject","url":"https://daily.dev/apple-touch-icon.png","width":180,"height":180},"sameAs":["https://twitter.com/dailydotdev","https://github.com/dailydotdev","https://www.linkedin.com/company/daily-dev-ltd"]},{"@type":"WebSite","@id":"https://daily.dev/#website","url":"https://daily.dev","name":"daily.dev","publisher":{"@id":"https://daily.dev/#organization"},"potentialAction":{"@type":"SearchAction","target":{"@type":"EntryPoint","urlTemplate":"https://daily.dev/search?q={search_term_string}"},"query-input":"required name=search_term_string"}}]} {"@context":"https://schema.org","@type":"TechArticle","headline":"Computer-use models will agentify the web, not APIs — Dhruv Batra, Yutori","url":"https://daily.dev/posts/computer-use-models-will-agentify-the-web-not-apis-dhruv-batra-yutori-fnoivacum","mainEntityOfPage":{"@type":"WebPage","@id":"https://daily.dev/posts/computer-use-models-will-agentify-the-web-not-apis-dhruv-batra-yutori-fnoivacum"},"datePublished":"2026-08-14T14:02:47.591Z","dateModified":"2026-08-14T14:28:39.369Z","description":"A talk from Dhruv Batra of Yutori argues against the popular idea that AI agents will interact with the web primarily through APIs (MCP servers, payment...","image":"https://i.ytimg.com/vi/Ki980nV0__0/sddefault.jpg","thumbnailUrl":"https://i.ytimg.com/vi/Ki980nV0__0/sddefault.jpg","isAccessibleForFree":true,"articleSection":"AI Engineer","inLanguage":"en","publisher":{"@type":"Organization","name":"daily.dev","url":"https://daily.dev","logo":{"@type":"ImageObject","url":"https://daily.dev/apple-touch-icon.png","width":180,"height":180}},"author":{"@type":"Organization","name":"AI Engineer","logo":"https://media.daily.dev/image/upload/s--u5PucxNT--/f_auto/v1724338940/logos/aidotengineer","url":"https://daily.dev/sources/aidotengineer"},"commentCount":0,"discussionUrl":"https://daily.dev/posts/computer-use-models-will-agentify-the-web-not-apis-dhruv-batra-yutori-fnoivacum","interactionStatistic":[{"@type":"InteractionCounter","interactionType":{"@type":"LikeAction"},"userInteractionCount":0},{"@type":"InteractionCounter","interactionType":{"@type":"CommentAction"},"userInteractionCount":0}],"keywords":"automation,ai-agents,mcp,crawling","timeRequired":"PT21M","video":{"@type":"VideoObject","name":"Computer-use models will agentify the web, not APIs — Dhruv Batra, Yutori","description":"A talk from Dhruv Batra of Yutori argues against the popular idea that AI agents will interact with the web primarily through APIs (MCP servers, payment...","thumbnailUrl":"https://i.ytimg.com/vi/Ki980nV0__0/sddefault.jpg","uploadDate":"2026-08-14T14:02:47.591Z","duration":"PT21M","url":"https://api.daily.dev/r/FNOIVaCum","embedUrl":"https://www.youtube.com/embed/Ki980nV0__0"}} {"@context":"https://schema.org","@type":"BreadcrumbList","itemListElement":[{"@type":"ListItem","position":1,"name":"Home","item":"https://daily.dev"},{"@type":"ListItem","position":2,"name":"AI Engineer","item":"https://daily.dev/sources/aidotengineer"},{"@type":"ListItem","position":3,"name":"Computer-use models will agentify the web, not APIs — Dhruv Batra, Yutori"}]} {"@context":"https://schema.org","@type":"FAQPage","@id":"https://daily.dev/posts/computer-use-models-will-agentify-the-web-not-apis-dhruv-batra-yutori-fnoivacum#faq","mainEntity":[{"@type":"Question","name":"Why can't AI agents just use APIs or MCP servers to access most websites?","acceptedAnswer":{"@type":"Answer","text":"Most of the web, especially the long tail beyond popular sites, will never expose APIs because the underlying institutions change infrastructure very slowly. Examples include restaurant menus stored as scanned PDFs or JPEG galleries, and school districts requiring Freedom of Information Act requests just to retrieve procurement documents by email and Google Drive. Only the most popular head-of-distribution websites are likely to offer real APIs. Developers weighing agent architecture choices can follow this API-versus-computer-use debate on daily.dev."}},{"@type":"Question","name":"Why can't a coding agent just read a website's HTML to get information instead of using vision?","acceptedAnswer":{"@type":"Answer","text":"Reading raw HTML often fails because much of what's displayed on a page, like stock status or a live sports score, is not present as static text but is rendered dynamically from a JSON response fetched asynchronously after page load. For example, an in-stock/out-of-stock indicator may only exist as a numeric quantity variable with a separate script controlling grayed-out styling, with no literal 'sold out' text anywhere in the HTML. Teams debugging why scrapers miss dynamic content can track these browser-rendering gotchas on daily.dev."}},{"@type":"Question","name":"How much does it cost to run a computer-use agent task compared to large frontier models like GPT-5.5 or Opus 4.7?","acceptedAnswer":{"@type":"Answer","text":"A smaller, specialized computer-use model called Navigator N-1.5 costs about 80 cents per task for roughly 20-30 interaction steps, compared to about $230 for the same task using trillion-parameter-class frontier models. Navigator N-1.5 also scored 97% human-eval accuracy on the Online Mind2Web benchmark, with only 8 of 300 trajectories judged incorrect, while offering lower per-step latency than larger models. Anyone comparing agent runtime costs across models can keep tabs on these benchmarks via daily.dev."}}]} ```