The SDLC is turning into an ADLC. Here's what changes.

The traditional software development life cycle (SDLC) – which typically includes phases like requirements, design, build, test, deploy, and maintain – is evolving into what some call an AI development life cycle (ADLC), according to a recent opinion piece. The core reasoning behind this shift is that the phases that once consumed the most time and resources in the SDLC are now becoming cheaper to perform, while the faster, cheaper phases are becoming more critical and time-consuming.

In this new ADLC model, the "building" phase – which in the SDLC was the most resource-intensive – has been largely automated by AI. Writing features now takes mere minutes for an AI agent, while tasks like context engineering, direction, generation, verification, and continuous harvesting have taken their place.

Context engineering now replaces the traditional requirements gathering phase. Instead of simply gathering requirements, engineers must create detailed "instruction files," decision records, and task files that the AI model reads and uses to generate code. In this sense, context has become the new deliverable, with the AI model relying entirely on this context to function.

Direction is now the human's role – the engineer selects options among those generated by the AI, rather than creating designs from scratch. Generation replaces the original build phase, as the AI quickly creates code which then requires verification. Verification now plays a larger role, as tests are no longer just checking the correctness of the code written by a human but rather ensuring that the AI-generated code meets strict contracts and specifications.

Continuous harvesting is a new phase where every session ends with the capture of what was learned into files that the next session can use.

While the deploy and operate stages remain largely the same, where generated code meets reality, the shift in engineering skills is notable. Writing precise constraints, judgment, and taste become more valuable than speed. Verification design, or knowing what proves correctness, has become a core engineering skill, not just a Quality Assurance (QA) task. Systems thinking is also emphasized, as the AI is very good at local tasks but lacks an understanding of the overall system.

The author acknowledges that this perspective is based on a single project, a solo developer working with an AI tool. On a larger team with existing legacy code, the ADLC might look very different. Also, as AI models improve rapidly, the verification burdens might decrease over time. However, the author believes that engineering will shift towards trusting AI-generated code, and this is a process problem rather than a model problem.

The teams that will succeed are those whose AI writes trustworthy code, which involves process improvements rather than just relying on better models.