In the AI Age

I spent a month trying to replace our coding models

I tried open-weight models in our development pipeline for a month. Reasoning limits, response formats and provider behavior made the switch more work than I expected. These are notes from June, without a cost benchmark.

One command gave my coding agent a project it could navigate

The OwlMeans Common scaffold starts with three TypeScript workspaces and package-specific agent guidance. A persistent feature still needs database configuration, validation and permission tests.

Sign-in was only the first part of the identity work

In June we built identity into generated apps: passwordless sign-in, an OIDC provider and permission handling. The work also forced us to define whose users and resources each application could access.

Agent as a project

How I organize AI colleagues as projects, with a defined role, linked memory, reusable skills and recurring work that can move between machines.

I stopped building features to sort out how the company worked

For two months I worked on the company around the product: a shared wiki, reusable workflows and a connection to the code. The result reduced repeated explanations, but it also gave me another system to maintain.

I had three spice racks and two coding agents

I moved our coding-agent guidance into one .agents folder. The harder part was keeping the skills current and stopping memory from becoming a diary.

A 2,000-year-old river gave me a better timeout

Our slowest test waits for a real coding agent inside a Kubernetes sandbox. Watching token usage helped detect a stalled run; isolation and cleanup handled the failures a longer timeout could not.

Preview was never the product

A preview is a useful place to build. Production configuration, a domain and a GitHub handoff need their own work. This is the design we started in June, with its October verification limits made explicit.