Since coding agents produce huge numbers of fine-grained tests, some demand that we give up unit tests. But the problem is not unit tests, it is units that are defined too small: tests that lock in the current class structure make refactoring harder despite high coverage and a high mutation score. As a measure of refactoring safety, both metrics are incomplete without the tested module depth. The article proposes testing domain components through their public API, determining a module-related mutation score and defining test boundaries as part of the architecture.
My idea of trunk-based development was naive: every commit straight to main, branches frowned upon. Research showed that short-lived branches with review are the norm and release branches are explicitly provided for. This self-test asks twenty questions about what actually happens in a team’s daily work and rates, following trunkbaseddevelopment.com, how far the practices of trunk-based development are in place: branch lifetime, integration, reviews, unfinished features, CI, local build, red trunk, code freeze, code ownership, story size and release readiness. The result is a degree of adoption with hints at possible bottlenecks, not a grade; questions and weighting reflect my current state of research.
Agentic AI can do more than develop new features: it can continuously improve the existing software. Three autonomous feedback loops show how this can work for correctness, performance and architecture, why the effort AI agents spend on changes could become a new maintainability metric, and why humans still have to set goals, limits and acceptable risks.
A guided rocket doesn’t have to be perfectly aligned at launch, as long as it knows its target and can correct its course. The same applies to AI agents: what moves to the center is the ability to aim at the right targets, check the direction automatically and readjust. Software development offers particularly good conditions for this.
AI software factories come as light factories with a human in the loop and as dark factories without human reviews. A gray software factory works autonomously by default, but decides for every single change, based on an automated risk and quality classification, whether a human needs to be involved.