Search across Doctrine, Practice, Journal, and Community
10 posts
Part three. We did not invent a new SDLC for agents. We put them into the Scrum most teams run, and kept the record: 311 build sessions, $3,268 at list price, 1,470 test runs judged on evidence, and the three things the record caught us getting wrong.
A demo is a foal. It cost nothing and it cannot finish a race. We are a very small shop that breeds the whole horse: data model, application, billing, infrastructure, the console you watch it from, and the support that runs after we leave. Here is the stable, and here is what is still dashed.
Part two. One real sprint meeting, kept as a model: 94 beats, 43 defined topics, 94 cited facts, a nine-step path, a self-test with no model call, the close against its pre-read, a $1.34 receipt, and all of it in Spanish.
Our AI agents sit in the sprint's meetings, and any claim that does not cite a fact from the frozen snapshot is struck before it is said. Nineteen minutes of the whole lifecycle on real screens, a real refinement that cost $1.34 and fell short of its own pre-read, and what does not work yet.
We pointed StellarView at Conduit, the most-implemented demo application on the internet, and asked for it as an operated platform rather than an app. It produced ten phases, 106 work units, and 357 questions it would not guess at.
A plan for a ten-phase platform produced 357 questions and refused to answer 281 of them. One had opened the repository, noticed docker-compose declared a build for a service whose Dockerfile did not exist, and asked who owned it. I had committed that file an hour before and left it unfinished.
Twelve hours and eighteen minutes, unattended, including two billing stops and a restart. 268 files, 25,857 lines, ten pull requests. CI green through phase eight and red on nine and ten, which we are leaving in.
Phase five wrote its own Terraform. Sixty five resources stand in a real AWS account behind a real certificate, and the application returns 503. The reason is a variable whose own description explains, in detail, why its default cannot work.
Monaco's John Tracy built a manager agent that spins off intern agents, each with seeded, production-like data. He is right. Here is how we fly it: parallel where parallel makes us accurate, single-agent and sequential where correctness is the whole job.
The old roles are collapsing. BA, architect, dev, QA, ops: they are becoming one person. Here are the skills that person needs, and how StellarView makes each one practicable.
