2026
-
When AI researchers compare notes too soon
A shared lead can help a team reason—and quietly decide what everyone investigates next.
-
Where should an agent’s tool call end?
A single call can contain a search, a program or another agent. The boundary determines which decisions return to the model.
-
Cheap to continue, hard to start
How cached context can turn an AI serving optimization into a choice about whose work gets admitted.
-
What an AI restoration asks us to believe
A reconstruction can correct a misconception while choosing details the evidence cannot settle.
-
Your AI should be a difficult customer
What an agent learns from a purchase—and what happens when a sales pitch becomes a preference it thinks you expressed.
-
Evidence, authority, and the cost of another question
How I separate evidence, authority, and spending from an agent’s choice of what to investigate.
-
Designing AI agents for decisions that keep changing
Lessons from trying to work out whether rebuilding an exotic could ever make sense for me.