Blog

Notes from the lab

Writing on long-running agent work, provenance, and the systems we build with both.

What a session leaves behind

The reasoning that produced yesterday's code is gone by morning. On the four kinds of record a resumed agent session actually needs, and the tool we built to keep them.

Read post

Narrow claims survive review

There is a sentence we keep deleting from our research notes. On null controls, ablations, and why we publish the regime where our own effect reverses.

Read post