3-Tier Web Access for Agents: Search/Fetch/Browser Separation Cuts Context Pollution by 82%
Dan-Mercede · reddit · 2026-07-06
The author proposes splitting Agent web access into three distinct roles: Search returns only candidate URLs without reading full text; Fetch converts pages into clean Markdown before passing them to the model (tests show a reduction from 9541 tokens to 1678 tokens on the same page, saving 82%); and Browser specifically handles operations requiring stateful interaction. The core philosophy is to prevent raw HTML—along with irrelevant navigation bars, scripts, and ads—from polluting the Agent's working memory. This layered design balances token costs with context quality.
More from coding & agent
- Hugging Face wraps 6-month Training Agents series with open recordings and code — SergioPaniego · 2026-10-26
- Agent verified 'success' after double-voting: a lesson in harness vs model bugs — TejasKumar_ · 2026-10-26
- diginomica: agentic automation hype is overdone, and 'The Hackening' of enterprise security has begun — jonerp · 2026-10-26
- 'Vibe verification': ~60% of AI-written formal specs prove the wrong property — basedjensen · 2026-10-26
- Devin launches Security suite powered by Claude Mythos with $100M+ credits — DevinAI · 2026-10-26
- HackBench injects reward hacks into agent trajectories to stress-test monitors — DanielKhashabi · 2026-10-26