Dorra Bartaguiz starts from AI's promise: more than 80% of developers believe it has increased their individual productivity, according to the 2024 and 2025 DORA reports. But speed means risk. Development time has shrunk, and the bottleneck remains code review. She notes that reading vastly outweighs writing: a ratio of more than ten to one according to Robert C. Martin, with about 70% of time spent understanding code versus 5% editing it, according to a 2015 study. AI accelerates syntactic generation and increases the review burden.
To find out whether generated tests can be trusted, she takes a real case: calculating the invitation date for a medical screening from the screening type, date of birth, and current date, with rescheduling rules for invalid, past, public-holiday, or weekend dates. Generated from a single prompt, the code is reported as 100% covered. She measures 33% overall coverage, 95% on the domain, flags a series of defects (hardcoded configuration, static methods, violation of the single responsibility principle, silent failures, magic values, tight coupling), and obtains a mutation testing score of 13.82%. Tests generated without rigor validate the AI's errors.
She then defines the harness as everything added around the model to improve generation; an agent is a model plus a harness. To make the agent more deterministic, she proposes craft practices and shift-left. The same case is revisited step by step: asking for domain definitions, breaking it down into rules, identifying examples and edge cases, writing Given/When/Then scenarios, generating acceptance and unit tests, then requesting a refactoring toward clean code, hexagonal architecture, and DDD tactical patterns. With AI, TDD becomes a practice for controlling an unpredictable agent; she cites Kent Beck, for whom teams that know how to test will adapt better.
Baby steps are not enough: the harness must be worked on, particularly through skills. Popular frameworks, from Superpowers to Skraft at AXA, rely on these practices. Finally, a product's life does not stop at the development cycle: equipping only the development team creates a new bottleneck. She proposes instrumenting the entire product cycle with specialized agents and guardrails against infinite loops, and sums up software factories as the combination of loop engineering, harness engineering, and human supervision.