Now
Last updated:
Atlas
Volume 1 (four chapters) is out in digital and paperback. Currently drafting Chapter 5 — Multi-agent orchestration without deadlocks, the opener of Volume 2. This one's taken longer than the previous four combined because the patterns I want to recommend aren't settled yet. I'd rather ship something I still believe in next year than something I'd retract by Q3.
Nexus
v1 ships May 24, 2026. Currently wiring up the routing layer — the part that decides which of your AI tools handles a given request — and a clean protocol for them to hand off context. v2 follows on June 27, 2026, with the memory layer that makes the whole thing actually feel coherent across sessions.
GrowthForge AI
Darshan and I are taking on a small number of engagements this quarter — the kind where a team has an AI feature half-built and we step in to ship it properly. If that fits, the door is open through the work page.
Reading & learning
Going deeper on programmatic eval design. The honest version of this work isn't "did the agent give a good answer" — it's "would I trust this output to ship to a user without me reading it." Building the harness to answer that question consistently is most of what I think about this month.
This is a now page, a tradition started by Derek Sivers.