Built systems at Unit 8200 that had to hold up against attackers. Anchor Browser gives AI agents computers for the real web.
Session
Workshop
Date
Monday, November 30
Level
Intermediate
// AI Engineering Skill /
Building agentic systems
// 01 / Bio
Idan Raman is co-founder and CEO of Anchor Browser, which provides scalable computers to AI agents that need to work on real websites in production. Earlier, at Unit 8200, he built systems that had to work in environments actively trying to break them.
Anchor is backed by Blumberg Capital and Gradient. In October 2026 it launched Action Index, execution memory that lets agents replay proven web workflows as deterministic code, reaching state-of-the-art accuracy at 1/98th of the cost.
// 02 / The session
From inference to indexing: frontier-level web agents at 1/100th the cost
Abstract
Most web agents re-derive the same workflow on every run, paying for model calls at every step even when the task was solved an hour earlier. Idan shows how Anchor’s Action Index turns successful runs into execution memory: parameterized, deterministic actions keyed by domain and replayed as code over MCP or REST, with the model stepping in only where a page has changed.
The benchmark: on 300 Online-Mind2Web tasks, a small model with Action Index matched a frontier model’s 93% success rate at $0.016 per task instead of $1.57.
What you’ll learn
How to separate known work from novel work in a browser agent
Designing step-level fallbacks so the model runs only where a page changed
What execution memory does to cost and reliability
AI DEV · BY DEEPLEARNING.AI
Join us at AI Dev
Meet the engineers shipping AI to production, and leave ready to do it yourself.