Episode Summary
Executive Summary: Davey Parikh explains Utori’s vision for “ambient” web agents that monitor the internet in the background and act on behalf of users, starting with Scouts, a product that tracks changing information across the web and surfaces concise reports. The conversation covers Utori’s hybrid architecture, visual browser navigation model, multi-agent orchestration, evaluation strategy, and how the product may expand from read-only monitoring to actions like purchasing or workflow completion.
Main Topics: Utori’s vision for the future of web interaction (Priority: 5/5): Davey argues that web use will move up an abstraction layer: users will describe intent while background agents execute workflows proactively, rather than people clicking through websites directly. Scouts as the first product wedge (Priority: 5/5): Scouts monitors the web for events users care about—such as product availability, price drops, job listings, apartments, news, and competitive intelligence—delivering reports rather than full task completion. Hybrid architecture: APIs, MCP tools, and browser use (Priority: 5/5): Utori combines direct tool access for agent-friendly sources with remote browser navigation for the long tail of websites that require clicking forms and page interactions. Visual browser navigation model training (Priority: 5/5): The team trains a browser-use model on screenshots rather than DOM structure, arguing that visual grounding generalizes better across sites and handles tricky UI elements like date pickers more reliably. Multi-agent orchestration and scale (Priority: 4/5): Because the system has access to 80–90 tools, Utori uses hierarchical orchestration and sub-agents to avoid context blowups and to coordinate parallel web searches and workflows. Evaluation, RL, and product iteration (Priority: 4/5): Utori uses step-level, trajectory-level, and report-level evals, plus human review and user feedback, and is now moving from supervised fine-tuning and rejection sampling toward reinforcement learning. Product expansion beyond monitoring (Priority: 4/5): Scouts is intentionally narrow today, but Utori plans to extend it toward actions like purchasing, broader workflow completion, and integrations via APIs/webhooks.
Key Arguments: Web interaction will become agentic and proactive, with agents handling chores in the background instead of users manually navigating websites. A useful product must push both the model and the product experience at the same time; strong tech demos without reliable UX, or UX without model capability, will not work. Browser AI should not just be an AI layer inside today’s browsers; the interaction model itself should change, with agents operating out of the user’s way. Visual screenshots are a more general and reliable input than DOM parsing for browser automation because websites differ wildly in implementation even when they look similar. A horizontal architecture that can monitor anything on the web is more scalable than building one-off domain-specific automations. Scouts is valuable because many information-seeking tasks require lightweight forms, hidden states, or broad coverage across many sources, making simple alerts or naive scraping insufficient. The system needs multi-agent and hierarchical orchestration because giving one orchestrator too many tools at once causes context and reliability problems. Utori’s near-term strategy is to earn trust with narrowly scoped, reliable product use cases and then gradually expand capabilities and autonomy.
Data Points: Years in AI: ~20 years - Davey describes his career span in AI, starting with computer vision PhD work. Time at Meta: ~8 years - He worked at Meta in FAIR and GenAI before co-founding Utori. Founders’ shared work history: 19 of 20 years - Davey and Dhruv have worked at the same employer for nearly their entire relationship. Weekly dinner tradition: 6–7 years - The founders held a weekly dinner called “brainstorming” for years before starting the company. Tool count: 80–90 tools - Utori’s stack has access to roughly 80–90 tools/MCP-style integrations. Product availability: waitlist - Scouts is currently available behind a waitlist with cohort-based onboarding.
Pivotal Quotes: "We will no longer be interacting with the web in the same way that we do right now." — Davey Parikh: Opening vision statement about the future of agentic web interaction. "We shouldn’t even be in browsers the way they are today, looking at web pages the way in which we do today." — Davey Parikh: Explaining why Utori sees AI browsers as an insufficient endpoint. "The web page is rendered for human consumption. ... over time, we just realized that consuming these web pages in the same way that humans are by just looking at visual screenshots is way more reliable and way more general." — Davey Parikh: Justifying Utori’s screenshot-based browser navigation approach.
Implications: The transcript suggests the first durable AI web products will be narrow, reliable, and background-oriented. For the industry, the winning stack may combine visual browser agents, tool use, and orchestration before expanding into full autonomous actions.