2. What changes when AI does the building · Convergence Programming

AI lets more people turn an idea into working software. The build can now start before anyone has written down all the details of the work.

A studio owner can describe a booking app and try it the same afternoon. Seeing it run helps her sharpen the idea: what should happen between sessions, who can move a booking, which rules change at another location. Her first prompt opens that conversation. The useful details emerge as the software meets the studio.

Software teams have long learned requirements this way. Developers interview users, test prototypes, revise specifications, and keep regression checks. AI can participate in all of that. The challenge becomes sharper when the person who knows the work cannot inspect the generated implementation, or when an agent changes more than that person can review. A plausible result needs an observable path from the owner's intent to the behavior she will rely on.

How this relates to tools you know

Convergence Programming proposes a common record for that path: what the build assumed, what it learned from the world, what each repair preserved, and which actions its evidence supports. Explore the neighboring approaches below. Their practices overlap, and each can contribute to this record.

The comparison covers eight overlapping approaches. Traditional software development can discover requirements through collaboration and iteration. End-user tools bring the person who knows the work into the build. Demonstration-based systems learn from examples and corrections. Program synthesis can refine candidates using counterexamples against a specification. Prompt-to-code tools can participate in longer review and testing loops. Tool-using agents can gather evidence from execution and the environment. Benchmarks evaluate the behaviors and conditions their tasks cover. Convergence Programming proposes evaluating the recorded trajectory across requirement discovery, repair, preservation, and permission to act. No approach is assigned a predetermined success or failure.

The overlap is explicit in the prior work. The Agile principles welcome changing requirements and collaboration with the people doing the work. Counterexample-guided synthesis refines candidate programs against a stated specification. Anthropic's agent-building guidance describes using tool results and code execution to ground progress in the environment.

Our proposed contribution is to connect these activities at the level of one build's recorded history, with a taxonomy of false convergence and an evaluation design for discovering hidden requirements. Better evidence can justify more useful action; unresolved requirements and changed conditions tell the builder where to look next.

For builders

Start with a decision your app makes. Record the requirement behind it, where that requirement came from, and a way to observe the behavior. Keep the check when you add the next feature. An existing development or agent workflow can already do this; Convergence Programming asks whether the accumulated record supports the actions you are about to enable.

For researchers

The proposed distinction concerns the unit of analysis and the obligations placed on evidence. It does not establish superiority over iterative development, synthesis, or agent architectures. An informative evaluation would compare implementations with comparable models, tools, evidence access, and oversight budgets, including strong baselines that discover requirements and preserve regressions. See the paper record's comparison for the scope of the claim.