Guides
How agents take in outside content, what the research shows about context, and the measurements behind it.
Four fetches, two refused
One AI agent, four web fetches, read from the record they left. Two pages were admitted, two were refused, and every decision has a reason a person can check.
How agents use outside content
How to assess permissions, source quality and embedded instructions when agents read external content.
Context window optimisation: a working guide
A guide to selecting, transforming and measuring the content an AI model reads, from source retrieval to answer quality and cost.
The state of the evidence, August 2026
A review of the research behind context optimisation, with evidence grades, replication status and limits for each claim.
Measurements behind the guide
Methods and reproducibility limits for token density, raw HTML overhead and overlap between search providers.