The problem was well defined.
The objective was execution. A 6DOF flight dynamics simulation, deployed in the cloud, needed to reach a level of fidelity and iteration speed that made it usable for real engineering decisions. The constraint was not ambition, but mechanics: correctness was costly, iteration was slow, and even small changes required careful verification across tightly coupled systems.
AI, at that point, was not positioned as a solution. It was a potential accelerant, useful for snippets, explanations, or isolated tasks, but not something you trusted with a system that needed to converge.
That assumption did not hold.
From Assistance to Coordination
The inflection point came from workflow design rather than model capability alone.
Following guidance similar to what Anthropic described around agents communicating via scratchpads, the system was structured in a way that traditional tooling did not support. Two Claude instances operated in parallel roles, each with a clear responsibility:
- One focused on reasoning and formulation
- The other on implementation, testing, and verification
They did not share a single conversational context. Instead, they communicated through structured scratchpads, external artifacts that forced explicit reasoning, intermediate results, and assumptions to be written down. This changed the dynamic: determinism emerged as assumptions became explicit, and reliability followed once failures were observable and repeatable rather than hidden in context.
Incremental Truth, Committed to Git
What followed was not a demo, but a process.
The simulation emerged through iterative refinement of a working baseline. Changes were introduced deliberately, tested continuously, and committed incrementally with explicit intent. Review cycles began to resemble human peer review more than prompt-response behavior. Each step produced an auditable artifact. Each decision left a trail.
This mattered. The simulation did not arrive fully formed; it converged. Convergence, not speed, was the signal.
Why This Was Different
The novelty was not that the AI wrote code. That had been possible for some time.
The novelty was that coordination replaced prompting.
Instead of asking a model to generate an end-to-end solution, the work was decomposed so that reasoning, execution, and verification were separated, externalized, and forced to reconcile with each other. In that structure, the AI was no longer a tool issuing answers. It became a participant in a constrained process.
That process held up.
By the end of the exercise, the result was not just a working simulation. It was evidence that long, multi-step, technically rigorous work could be sustained, provided the workflow was designed to support it.
That constraint shaped the work that followed.
Subscribe for new essays
Get updates when new writing goes live.
