Our system identifies the smallest pieces of code required in order to make the requested change to the system; this mitigates the "context window" question for ongoing product enhancement. We also have specialized techniques to minimize the amount of context we need in any given request — precision code snippets (where we provide code with line numbers surrounding the actual target), codebase chunking strategies, compression algorithms, etc. This mitigates issues where models are limited more on their output windows than input windows.The largest codebase we have reliably worked with thus far is 5 million lines, and there is no theoretical upper bound.

A big part of Stride 100x is decomposing huge, even intractable problems into pieces that LLMs can solve, in ways humans can understand. We do much of this through workflow design and prompt engineering, but have also built specific capabilities into our tools (precision code snippets, dependency mapping, codebase chunking) that help us handle extremely large problems. See our case studies to see how we handle hundreds of files, thousands of errors, millions of lines of code, etc.

100x is great at refactoring — just provide your guidelines and suggested architecture if you have it, and we can deliver large scale refactors that meet your goals. This isn’t as simple as pressing a button — no large refactor can happen without humans in the loop and agreement on target architectures and patterns — but we have proven that we can execute on big refactors with a combination of AI tools and smart humans in a fraction of the time you’d think.Yes, our tools document existing code! Test generation and documentation are extremely similar. We can do either one, or both, using our Conductor and 100x tools. We have also done a lot of interesting work with sequence and dependency diagrams, which can be a big part of the architecture alignment process where we design a large-scale refactor for human approval.

Stride and 100x are language agnostic, we can and do work with anything — and our experience building across a wide range of languages helps us configure the LLM agents more quickly and efficiently than less experienced operators. The point about mainstream languages is relevant (the more examples the better) but we are also providing your code and team guidelines as context, so that also gives a good pattern to follow. Generally anything modern will be fine.

Stride’s code generation tools leverage proprietary methods for getting consistent, usable code from LLMs; this represents significant R&D investments in agent configuration and optimization, as well as a commitment to continued development, iteration, support and enhancement. We are also solving problems at scale; most of the solutions you’re reading about are proofs of concept, not enterprise grade. Essentially, we see this as our core business – you already have one of your own!