Notes
Notes
What broke, what was measured, and what the measurement turned out to mean. Written while building a search API for agents.
-
A web page is mostly not the article
We measured what survives extraction across 74 real pages our search API served: 4.2% of the bytes that came off the wire. What the other 95.8% is, and why it matters when you are paying per token.
-
An empty result is not an error
The most common bug in agent search integrations is treating an empty result list as a failure, or a failure as an empty list. Why status codes have to carry that difference, and what happens when they do not.
-
What it costs to give an agent web access
Most agent searches do not need the page. A four-rung depth dial, what each rung actually buys, and why the cheapest useful answer is usually the snippet you already have.
-
A search API is a prompt injection surface
An attacker targeting your agent through search does not need to reach you. They only need to rank for a query you run. What that changes, and the three defences that survive contact with a fully persuaded model.