It's still either Scrum, Waterfall or cowboy in the age of AI agentic programming

Anybody can now have their own dedicated development team for $100 a month, and can do whatever they dream up. Everybody can be a Product Owner.

Just one small caveat: this is like spawning a million Product Owners with strong product vision but zero Scrum training, and giving them a development team with strong development skills and zero Scrum training too. This is a recipe for total chaos from a productivity, quality and scaling point of view. You know what you want to build (high level), you know how to build it (at the unit/component level), but you have absolutely no clue how to orchestrate the whole thing. That's exactly what Scrum is for, and you normally need Scrum Master skills to get it.

But most people don't have it. So just as we already have a term for AI generated social media garbage (AI slop), we now have tons of "AI slop code". Call it the Spider Man side-effect: with great power comes great responsibility, and very few people truly know Scrum, so they are riddled with misconceptions, misunderstandings and frustrations. But Scrum continues to be Scrum: the uncontested, #1 way to build software products in the world. There simply is no alternative. What do you think happens when you give a Product Owner unmetered access to a team that's eager to code? Cowboy. Or maybe Waterfall, if the spec is detailed enough.

Anthropic and many other people recognized this and set out to tackle the challenge: agents and agentic teams need a development methodology just like humans do. Skills and techniques for managing work started appearing. One of the first was Geoffrey Huntley's "Wiggum loop", a simple idea of iterating over something until it reaches a state of completion, some sort of primitive shell of Scrum, where the iterations are machine-defined and the point of completion is blurrily defined. It instantly got misinterpreted by Anthropic, who shipped a plugin that did not reset the context on each loop, which earned it plenty of mockery (although I suspect their engineers were betting on automated context compaction). We also saw ambitious projects like Steve Yegge's Gas Town emerge, and skill combos like BMAD. Many of these look a lot like Agile practices and Scrum ideas wrapped in some extra thoughts by the authors. Eventually Jesse Vincent's Superpowers landed in Anthropic's official plugin marketplace: a fine-tuned set of skills for Claude Code that clearly and explicitly brought several agile practices (mostly technical) into a neat set of agentic guidelines.

But none of them address the problems Scrum is designed to solve: building a product with high quality, iteratively and incrementally. Several actually drive you in the opposite direction, back to Waterfall. "Build the right spec up front, then give it to an army of agents" is the new mantra. But is this the right approach? Garbage in, garbage out, right? You cannot AI your way around the complexity problem: until you build and test a product with real end users, you don't know what you really needed to build. That's why we do Scrum.

So what would it mean for an agentic team to do Scrum?

• The human is the Product Owner
• The agent is the Development Team
• The skills and AGENTS.md are the Scrum Master
  

The third is the key insight. In Scrum, the Scrum Master job largely consists of "herding cats": trying to get people to define and respect working agreements. A Scrum Master's day goes into supporting the PO in maintaining the Product Backlog and breaking down PBIs into smaller pieces, helping the developers refine and estimate, and promoting technical excellence.

Most, if not all, of this can now be handed to AI agents. So the Scrum Master job becomes at least partially automated. Not entirely, but a large part of it. The key is you do need a solid Sprint Zero where you define and set up all the working agreements, tooling, environments, etc.

Which sounds like it puts us back where we started, since the only people left are Product people, not Agile Coaches. They can't write working agreements they've never experienced. But the good news is they don't have to. Somebody else can writes them for them once, and it's a one-time cost, not a per-project one. You start from a known-good set and change only what your context actually demands.

For example, on a new product we're developing, we run a strict agentic Agile methodology that closely follows what we would do with a human team:

• We create and split PBIs in a Product Backlog, with a Definition of Ready and a Definition of Done
• We do refinement and "PBI Planning" (the equivalent of Sprint Planning, but on a PBI-per-PBI basis - something a bit Scrumbanish)
• We implement the technical practices that keep quality high: TDD, code reviews, continuous integration, devops delivery pipeline (automated acceptance testing, automated deploys, database migrations)
• We update documentation as part of the DoD
  

Agents are very good at following instructions if their context is well structured and not overloaded. Ours always follow the methodology, which means we have few regressions and are able to maintain a constant velocity. In my experience, a typical User Story takes about 20 minutes to go through the full lifecycle from not ready to deployed into production, on a codebase of around 800,000 lines.

So: it's either Scrum or cowboy. The good news is that most of good Scrum can now be done for you automatically if you just figure out the right skills and CLAUDE.md. The bad news is that being a Product Owner still can't. That part you still have to learn. That, and how to spawn and control the AI agent developer army. We can help you with all of this.