The Agentic Web is Next: Building the Infrastructure for AI Agents on the Web | RAISE Summit 2026

Company Overview & Strategic Vision

• Yotori, founded by Abhishek (Co-Founder/Co-CEO), operates a 14-person team based in San Francisco focused on web agents—AI agents that take actions and interact with websites via APIs for enterprises.
• The company identifies a market inflection point where web interaction must shift from manual human clicking to autonomous agent-driven actions over the next 2–10 years.
• Yotori addresses critical web automation challenges including the lack of APIs for the "long tail" of websites, aggressive bot detection, authentication walls, irreversible actions (e.g., incorrect purchases), and the high cost of frontier models.
• The core thesis posits that the next paradigm shift follows coding agents, targeting the full spectrum of "labor on the web" such as form submissions, transactions, CRM management, and cross-site orchestration.

Technology Stack & Model Performance

• The core component is "Navigator," an in-house post-trained computer-use model; the current public version is Navigator N1.5.
• Navigator N1.5 is described as "pointer dominant" in accuracy, cost, and latency compared to frontier models, specifically surpassing Opus 4.7 and GPT 5.5.
• Benchmark performance: Navigator N1.5 ranks top of the "mind2web" academic leaderboard (human-evaluated) and outperforms competitors on the NaviBench and Westworld benchmarks.
• Model capabilities include a hybrid toolset allowing the agent to either predict UI clicks/typing or write JavaScript to directly manipulate the DOM, reducing task steps (e.g., extracting product variants) to as few as five steps.
• Cost and latency advantages: Navigator N1.5 is 2x to 5x faster and cheaper than frontier models, fundamentally altering the unit economics of product features like form filling or data extraction.
• Training methodology focuses on post-training (fine-tuning) open-source models (e.g., from Quen) using mid-training, supervised fine-tuning, rejection sampling, and reinforcement learning against programmatic or human-written verifiers.
• A significant portion of reinforcement learning data is collected via real-world rollouts where agents interact with actual websites, supplemented by sandbox environments for high-stakes tasks.

Product Architecture & Deployment

• The stack comprises the Navigator model, a cloud browser fleet for parallel task execution, and an agentic harness for integrating APIs and MCPs.
• Three API abstraction levels are offered:
  • Computer Use Model: Token-in, token-out interface accessible via sign-up.
  • Browsing API: Bundles the model with browser infrastructure for developers avoiding local setup.
  • Research & Scouting API: Enables seamless mixing of browsing tasks with other external APIs and MCPs.
• Production scale: The platform is handling over 100 million API calls per month.
• Deployment channels include first-party endpoints (powered by Crusoe, BaseStand, Together), managed inference marketplaces (Crusoe, Foundry), and integrations with cloud browser providers.

Real-World Enterprise Use Cases

• Fortune 100 Advertising Company: Validates promo code availability for 100,000+ merchant websites by programmatically navigating checkout flows to extract structured data, a task impossible via standard APIs.
• Indian Fintech Giant: Uses a single N1.5 model for end-to-end agentic checkouts to perform fraud and risk detection by scanning high-risk merchant sites for specific mentions during the transaction journey.
• Global Browser Provider: Integrates Yotori's API into a side-pane feature allowing consumers to prompt an agent to complete tasks across any website within their own browser.
• Diverse Bottom-Up Adoption: Use cases range from shopping on eBay, extracting government data, managing Stripe dashboards, to frontend QA testing for "vibe coding" products.

Future Outlook & Roadmap

• Navigator N2: Scheduled for release in 1–2 months, expanding capabilities beyond browsing to broader general computer use.
• Market Strategy: Advocates for a heterogeneous, multi-model future where domain-specific constraints (latency, cost, integration) dictate model selection rather than using generic frontier models for all tasks.
• Call to Action: Interested in engaging with businesses dealing with heavy browser workflows (data extraction, form filling, checkouts) for potential partnerships on N2.
• Current Availability: APIs are accessible via the website, with supporting tools including MCPs, CLIs, and a Chrome extension.