Build Stories & Field Notes
New pieces appear here as they are published. Start with the newest article, browse a hand-picked featured piece, then read the rest in reverse release order.
Newest article
For a while, I thought memory was one giant problem. If the assistant could retrieve something useful and sound informed, I counted that as progress. That standard breaks down once continuity becomes operational. Memory stops being one job and becomes several different jobs that need to be separated if you want the architecture to stay understandable.
Featured pick
The first version of my assistant stack looked more impressive than it really was. On paper, it had everything that makes agent work feel futuristic. There was a primary assistant for direct interaction, sibling agents with different roles, Telegram routing for real-time use, memory files for continuity, and a growing set of automations that made the whole thing seem alive. In screenshots and short demos, it looked l
The first few days with an agent almost always feel better than the next few weeks.
There is a stage in almost every builder's path where the work is genuinely interesting but not yet commercially legible.
The first version of the sales agent was polite, clear, and technically competent.
We did not add Postgres because Chroma was broken.
Most people imagine agent failure as something obvious.