Introduces the 'coding agent sandwich' pattern for building GenAI-based development workflows. The pattern layers plain old code (POC) as the top orchestrator, LLM invocations as the filling for non-deterministic tasks, and deterministic tools as the bottom slice. Illustrated through the `implement plan` workflow in the open-source `idea-to-code` project, which uses TDD to turn a generated plan into Git commits. The core principle: use deterministic code whenever possible and invoke LLMs only for tasks that genuinely require probabilistic reasoning, such as generating commit messages, fixing CI failures, or responding to PR comments. The architecture is described as fractal — tools in the bottom slice can themselves be sandwiches if they incorporate LLM calls.
Table of contents
An overview of the implement plan workflow §Why not implement the implement plan workflow using a coding agent? §The top slice: orchestrating the implement plan workflow using plain old code §The filling: invoking the coding agent for narrowly defined actions §The bottom slice: deterministic tools used by the coding agent §Functional decomposition, layered and fractal §Use POC whenever you can, LLMs only when you must §Need help with modernizing your architecture? §Sort: