Philippe Martin presents the BMAD method (Breakthrough Method for Agile AI-Driven Development) as an essential "urban planning blueprint" for structuring the use of agentic AI in the software development lifecycle, in contrast to "vibe coding", which generates explosive technical debt.
The problem: since the arrival of LLMs, many teams have been "coding by feel" with AI — prompting, retrieving code, manually adjusting, starting over. In the short term, this is exhilarating. In the long term, it is like letting each developer negotiate alone with a contractor: extensions everywhere, no overall plan, an electrical network ready to catch fire.
The BMAD response: organize agents IA as a disciplined construction crew. Each agent plays the role of a specialized foreman: one agent clarifies the requirement and produces a real specification rather than a vague prompt, an architect proposes options aligned with standards, "craftsmen" generate code, tests and documentation, and controllers track defects and regressions. The human remains chief architect and project owner: defining the vision, the non-negotiable constraints, and making the structuring decisions.
Integration into the SDLC: the author proposes three insertion points. Upstream, agents turn a business intent into a structured PRD much faster. In the middle, they propose architectures and generate code within a defined framework — "you don't ask the AI to 'write me an app', you ask it to 'implement this module within this precise design'". Downstream, agents monitor, analyze logs and feed a prioritized technical debt backlog.
Legacy and technical debt: BMAD pushes teams to address the structure rather than the paint layer. Agents map the code, identify at-risk areas and propose progressive refactorings. Systematic generation of tests and documentation helps pay down the backlog.
Evolution of roles: the autopilot analogy in aviation illustrates the transformation — the pilot has not disappeared, but the job has changed. Developers become solution designers and demanding reviewers, architects orchestrate a system in which enterprise constraints are injected into the agents, and managers learn to frame the problem and evaluate the value produced by the hybrid chain.
Recommendations: start with the rules of the game before the tools, choose concrete pilot use cases, measure and communicate the results, and accept that the framework will evolve — just as a good urban plan allows a city to be renovated without tearing everything down each time.