On the bench
The Real Problem With AI Agents Is Context Selection
select, don't shovel
Every agent failure I’ve debugged this year — mine and other people’s — has had the same root cause, and it wasn’t intelligence. It was context: the model either didn’t have the one fact it needed, or it had ten thousand facts it didn’t need and reacted to the wrong one.
We talk about models the way people talk about engines — more power, more capability. But a bigger engine doesn’t help if the fuel line feeds it the wrong mixture. Context selection is the fuel system of an agent, and almost nobody instruments it.
What this draft is collecting:
- Real failure cases from my diagnostic pipeline where retrieval picked the plausible-but-wrong KB chunk
- Why “stuff the whole context window” measurably degrades behaviour
- The front-controller pattern: a cheap model deciding what an expensive model gets to see
On the bench. Each claim here is getting a logged example before this ships.