The Double Diamond of AI Feature Development

The Double Diamond framework, born from design thinking, describes how creative work oscillates between divergent exploration and convergent execution. With AI agents in the development loop, this pattern becomes more pronounced—and more powerful.

The Classic Double Diamond

The British Design Council formalized the Double Diamond in 2005, but the pattern is ancient. First, you expand outward: explore the problem space, gather requirements, consider possibilities. Then you narrow: define the specific problem you're solving. That's the first diamond.

The second diamond mirrors the first but for solutions. Expand again: explore possible implementations, prototype ideas, experiment freely. Then narrow: select the best approach, refine it, ship it. Two diamonds, four phases: Discover, Define, Develop, Deliver.

AI Changes the Shape

Traditional development constrains how wide you can diverge. Exploring five implementation approaches takes time. Prototyping three different architectures is expensive. So teams converge quickly, often too quickly, settling on familiar patterns before fully exploring alternatives.

AI agents change the economics. When an agent can scaffold three different implementations in an hour instead of three days, the “expand” phase becomes dramatically wider. You can actually explore those five approaches. You can prototype all three architectures. The divergent phases of the diamond grow.

But here's what's subtle: the convergent phases also change. With more options to evaluate, selection becomes more important. Human judgment becomes the bottleneck—not in execution, but in deciding which execution to pursue.

The First Diamond: Problem Discovery

Discover (Diverge): AI agents excel at codebase archaeology. Point an agent at a repository and ask: “What are all the ways this feature could break?” or “What patterns exist in similar features?” The agent can read every file, trace every dependency, document every edge case. This discovery phase was historically limited by developer attention. Now it's limited only by token budget.

Define (Converge): With comprehensive discovery, definition becomes more precise. Instead of vague requirements like “add user authentication,” you can define specific scope: “Add session-based authentication using the existing UserService, with JWT tokens stored in HttpOnly cookies, compatible with the current middleware stack.” The more you discover, the better you define.

The Second Diamond: Solution Delivery

Develop (Diverge): This is where AI assistance shines brightest. Give a well-defined problem to multiple agents, or run multiple iterations with different constraints. Generate the solution using existing patterns. Generate it using a new library. Generate it with maximum type safety. Generate it prioritizing readability. Each variant is cheap to produce.

Deliver (Converge): From the generated variants, select and refine. This is the critical human function: not writing code, but evaluating code. Which implementation best fits the codebase? Which will be maintainable in six months? Which has the right tradeoffs? The agent produces options; the human selects direction.

Practical Application

When starting a feature, explicitly structure your work around the diamonds. First, diverge hard on discovery: have agents analyze the codebase, identify constraints, surface hidden dependencies. Then converge to a precise definition.

For implementation, diverge again: generate multiple approaches, even ones you think you won't use. The cost is low; the learning is high. Then converge to delivery: pick the best path, refine with human judgment, ship with confidence.

The discipline is in the transitions. Don't converge too early in discovery; you'll miss important constraints. Don't diverge too long in development; you'll generate options you never evaluate. And always remember which phase you're in—the goal of divergence is exploration, not production.

The Meta-Pattern

Perhaps the deepest insight is that AI development itself follows the double diamond. We're currently in a massive divergent phase: exploring what AI agents can do, experimenting with workflows, generating new patterns. Eventually, the field will converge on best practices, standard tooling, and established patterns.

Until then, embrace the expansion. The wider you diverge now, the better your convergence will be when the dust settles.