Non-Determinism Doesn't Matter

The piece argues that non-deterministic outputs from large language models are not a fatal flaw for software development because correctness requires a precise result, not a single fixed sequence of steps to reach it. It reframes LLMs as probabilistic actors that generate code as one of many possible means to an end, noting that programming has always tolerated multiple correct implementations across languages and stacks. Engineers' job becomes narrowing the model’s probability space using context engineering, multi-component prompting, skills, harnesses, graphs, and memories to produce reliably useful outputs even when some randomness remains.

The practical solution presented is to convert nondeterministic model output into deterministic tools that agents can call: tested, authorized functions that yield the same exact outcome on every execution. That pattern reduces token usage, creates reusable abstractions, and mirrors historical compounding automation in DevOps, but with AI building and invoking the abstractions at machine speed. Industry examples and personal production experience show this approach turns present randomness into future determinism by having agents create and rely on stable primitives rather than depending on perfectly repeatable raw model text.