Why I applied Agile to vibe coding
When I started working with AI, it felt effective and natural to apply a short-cycle approach: an initial phase similar to a Discovery Sprint, then the work broken down into smaller sprints, with a backlog holding the priorities together. I realized almost immediately that an AI agent — or several AI agents together — behave a bit like a multidisciplinary team: one writes, one reviews, one tests, each with its own role.
The difference from working alone is that here the actual production — writing the code — is delegated to the AI. The effort shifts elsewhere: to designing what's really needed and supervising what gets produced, rather than to execution itself.
In practice, each cycle more or less always took this shape:
In my case, the backlog is simply a markdown file with a list of checkboxes — nothing sophisticated. But it does exactly its job: keeping what comes next always clear, and making it possible to reorder priorities without losing the overall picture when the AI proposes several different paths for the same problem.
The test and UAT part remains the one that demands the most discipline: the generated code "looks" correct, it compiles, it does what it seemed it should do — but only by actually trying it, with real data and real edge cases, do the real problems surface. The feedback on those problems is what feeds into the next cycle, updating the backlog instead of moving forward on shaky ground.
There's something that became clear to me working this way, sprint after sprint, and it's still too undefined to deserve more than a mention: if an AI agent works well inside a well-scoped sprint taking on a specialist role, what happens if you have more than one working in parallel, each on its own slice of the backlog, with something above orchestrating who does what and in what order the results get integrated?
It's the idea of an application to govern vibe-coding development with several AI agents running concurrently — it isn't a project yet, it's more of an intuition that keeps coming back, born precisely out of treating AI as a multidisciplinary team instead of a single executor. If it takes a more precise shape, you'll probably find out on the next page of this logbook.
Applying this approach didn't make vibe coding simpler, but it gave me a ready-made vocabulary and framework, instead of having to invent a new one from scratch while I was also learning to work with AI.