experiments

structured run logs for multi-step agent work — what ran, what failed, what we kept.

small, boring fixtures that catch regressions before a demo does.

treat prompt changes like code: reviewable diffs, named versions, rollback.

caps on tool calls and tokens so agents stay useful under cost pressure.

book-jacket paper, iron gall ink, and calm type for builder-facing tools.

lightweight session memory that prefers notes over opaque embeddings.