← Spanwise

AI Development

Notes and essays on building consequential software with AI agents.

Rasti: How We Plan with Control Points

We rebuilt Matt Pocock's wayfinder pattern on our own substrate: maps of decision tickets living in the same database as the work, every resolution punched as a verbatim decision record. Named for the orienteering control point.

Read →
The Gate Walk: Deciding at the Speed Agents Ask

Agents hit judgment calls faster than anyone can answer them well. The gate walk saves them up under one honest tag, packages each with evidence and one recommendation, and records every answer before the next gate.

Read →
The Desk: Reviewing at the Speed Agents Write

Agents draft long documents faster than anyone can read them. The desk is the review loop: immutable versions, anchored comments, and an agent that reads, answers, and resolves markup from every reviewer.

Read →
The Layer That Remembers Why

Design rationale, revived, with AI carrying the half that killed it. Vedos records why every screen exists, so the layer that moves at the speed of taste finally has a memory.

Read →
The Layer That Can't Lie

Twelve agents restyled half our aviation product overnight and its judgment never flickered. On separating the layer that can't lie from the layer that's allowed to move at the speed of taste.

Read →
Sunday: a durable operating model for consequential agent work

When agent work gets big enough, you don't need a smarter conversation. You need a program: a coordinator, bounded workers, independent review, and a durable ledger that survives a crash.

Read →