Optimising the Strike Software Factory

The Context

Part II of the series that started with Building the Strike Software Factory. An agent is a model, harnessed, operating in an environment. The environment is the part we own.

Quality comes from two mechanisms with very different economics. Deterministic gates (compilers, analysers, linters, tests) are a fixed cost: build once, run free forever. Non-deterministic guidance (prompts, skills, review agents) is a marginal cost paid in tokens on every run. Prompting an agent not to do something is coercion. A build-breaking rule is enforcement.

Why This Matters

Human review time is the most expensive line item of all. Every mistake class moved into a deterministic gate - as close to the keystroke as possible - means PRs arrive clean and reviewers spend their attention on architecture rather than mechanics.

The post covers the six layers of the factory, using evals to regression-test skills, corpus hygiene, and why repository-level assets keep their value across model generations. A faster agent that creates twice as much review work is not an optimisation.