Перейти к содержимому

How AI Agents Can Act on the Web Without Getting Blocked | Andrey Marey

To Data & Beyond

0:00 / 0:00

How AI Agents Can Act on the Web Without Getting Blocked | Andrey Marey

204 просмотра · 12 дн. назад
To Data & Beyond
204 подписчика
204 просмотра · 12 дн. назад
Most websites block AI agents, and browser agents that drive a screen through screenshots are slow, expensive, and fail often. In this free To Data & Beyond live session, I talk with Andrey, co-founder of AgentWeb, about how they give agents structured, network-level access to sites like LinkedIn, Amazon, and DoorDash, so an agent such as Claude can read and act on your behalf without a cloud browser in the loop. Andrey walks through a real run where Claude pulled everyone who engaged with a founder's last six LinkedIn posts into a spreadsheet and sent her a connection request, for about $1.30 and roughly 11 minutes of work. We then go into the engineering behind it: session and token handling, CAPTCHAs and bot protection, residential IPs and geo consistency, safe throttling, indexing new sites on demand, and why AgentWeb is not a fork of Chromium. The session ends with a live setup of AgentWeb as a custom connector in Claude, a LinkedIn messaging demo, and a Q&A on guardrails, privacy, pricing, context-window cost, and payments. Timestamps 0:00 To Data & Beyond intro 0:03 Welcome and introducing Andrey from AgentWeb 0:23 The problem: sites are blocking AI agents 1:05 Why computer-use and browser agents are slow and expensive 1:36 Demo: LinkedIn post engagers to CSV plus a connection request 3:43 Residential IPs, session handling and safe throttling 4:38 Who uses it: individuals and personal-assistant companies 5:08 The engagers spreadsheet, cost and run time 6:52 Supported sites: 1,700 sites, 400K+ subdomains, 48K hotels 7:27 Indexing a new site on demand 8:42 Technical challenges: tokens, CAPTCHAs, bot protection and IP geography 11:20 API docs and the public sites index 12:06 Moving from manual indexing to autonomous indexing 14:07 AgentWeb vs Claude in Chrome and Chromium-based browser agents 16:11 Why screenshot-based agents burn tokens 17:32 Agent payments: single-use cards and cards on file 18:47 Guardrails: read vs write actions and user confirmation 22:13 Pricing: per-action billing and no charge for failed actions 24:09 Live setup: adding AgentWeb as a custom connector in Claude 28:27 Live demo: finding a LinkedIn thread and sending a message 30:53 The vision: agents that act on the web for you 31:56 Privacy: what data AgentWeb sees and what it never sees 33:29 Availability outside the US 34:23 Inside a response: structured data and the developer API 38:41 Context-window cost and how tools get selected 40:22 Free credits for attendees 42:19 Where AgentWeb is heading in the next 1 to 5 years 44:34 Wrap-up Links AgentWeb: https://agentweb.us