The week in AI, read with a data and retrieval lens. One email a week.
This week's engineering attacked the cost of putting text into a model and holding what it already read, which is where teams go once long context runs into either the model's limit or the invoice
The artifacts agents read to decide what is true, forum threads and benchmark tables and model pages, became the contested surface this week, and one of them turns out to be trivially cheap to write for someone else
The week's better work was about leaving a readable record of a system, because the corpus a project publishes about itself is what agents will learn it from
The week's proposals all moved checking after the fact, to sampled trajectories and runtime enforcement, because agent output now exceeds what anyone will read before it lands
Three document parsers, a recursive context library and a renaming of prompt injection all landed on the same layer, the intake path a model reads before it answers
Every gain worth noting this week was made outside the model, in skills, loops, replay and the index, which turns the parts a team owns into the parts that improve
