Questions I’m carrying

A few threads I’m actively collecting material for.

How much can an interface compress?

Verba compresses speech into text. Monarch compresses protocol state into lending decisions. Agent tools compress external signals into work. The shared problem is deciding what complexity to hide and what must remain visible.

What should agent work leave behind?

For autonomous work to be trusted, it needs residue: diffs, tests, URLs, traces, notes, or decisions that can be inspected outside the chat.

Can reading tools change what gets noticed?

ai-reader points at a problem I still like: not summarization as replacement for reading, but software that helps attention find structure without stealing the act of understanding.