Justin McCarthy (Diffusion founder) drops a critical insight on agent reliability: success isn't about smarter models, it's about context architecture.
The core problem: dump an agent into a human-oriented codebase and it'll load the wrong context every time. You can't even trace which files get pulled during context priming before the agent makes its move.
His solution: design environments where the right context is inevitable. Think of it as context determinism over probabilistic retrieval. Instead of hoping the agent finds the right files through RAG or embeddings, structure the codebase so the correct context is the only path forward.
This mirrors how compilers enforce type safety rather than hoping for correct runtime behavior. For agent systems, it means rethinking repo structure, dependency graphs, and how context windows get populated before any code generation happens.
The core problem: dump an agent into a human-oriented codebase and it'll load the wrong context every time. You can't even trace which files get pulled during context priming before the agent makes its move.
His solution: design environments where the right context is inevitable. Think of it as context determinism over probabilistic retrieval. Instead of hoping the agent finds the right files through RAG or embeddings, structure the codebase so the correct context is the only path forward.
This mirrors how compilers enforce type safety rather than hoping for correct runtime behavior. For agent systems, it means rethinking repo structure, dependency graphs, and how context windows get populated before any code generation happens.