Long form. Source-linked. No marketing.
Posts from inside the build of Team-X. Architecture decisions, role-spec patterns, post-mortems, and the rare manifesto. Every claim links to a primary source.
The privacy setting that enforced nothing
A privacy toggle that changes a label and nothing else is worse than no toggle, because the user stops looking. Team-X shipped one. v3.5.0 enforces the tier on every model call and deletes every control it could not back.
Read post →-
Zero native integrations. By design.
Integration count is the vanity metric of agent platforms. Every connector is a credential someone holds and code someone must patch. I moved that decision to the operator, and I will show you where the boundary leaks.
Read → -
3,640 lines of code nobody could reach
Compiling and passing tests is not done. Done means reachable from an entry point. Team-X carried 3,640 lines of renderer code that was not, including install dialogs that invented the permissions a user would approve.
Read → -
Our vector database never ran. Retrieval was fine.
The README promised sqlite-vec embeddings. The code path threw on every call and a catch block absorbed it. Retrieval worked anyway, which is exactly why nobody looked. Here is what the audit found and what I changed.
Read → -
Autonomy needs an off switch, and a ceiling
The market is racing to ship agents that act without asking. The hard product is the set of controls that makes it safe to let them. I audited my own v3.4.0 code, control by control, and some of it is thinner than the screenshots.
Read → -
Status is a word, not a color
A red, amber, green dot fails readers who cannot tell the colors apart, and WCAG says color cannot be the only cue. Here is the status system Team-X v3.4.0 uses instead, and the contrast bugs it shipped before fixing them.
Read → -
Local-first is an architecture, not a setting
Local mode is a toggle. Local-first is a decision about where the state of record lives and what the process may do by default. I walked the Team-X v3.2.1 stack to see which of my own claims survive a grep.
Read → -
Eight turns, 8,000 tokens, 120 seconds
Most agent pitches sell autonomy as run until done. A loop with no hard stop is a billing incident. Here are the four caps Team-X enforces at v3.2.1, what happens when one trips, and where the limits leak.
Read → -
A role is not a prompt
Most AI agents are one paragraph of stage direction. Team-X defines an agent as a schema-validated role spec, and the runtime enforces some of it. This is a field-by-field audit of which parts, at v3.2.1.
Read → -
Our docs described a product that did not exist
Everyone audits the chatbot that invents a fact. Nobody audits the manual. Team-X's docs described a hosted platform that was never built, and strangers could read every line of it. Here is what was wrong and what the repair missed.
Read → -
Run an AI company. Not a prompt.
Prompt engineering is a dead end. The interesting unit of work is no longer the prompt; it is the org. Here is the case for treating AI like a company, and the architecture that makes it real.
Read →