Developer proposes limiting AI agents to JSON output while local tools handle execution

A software developer has outlined a pipeline design where an AI language model is restricted to a single task: filling a predefined JSON contract, while all file edits, tests, and commands are handled by conventional local tools. The approach aims to prevent common agent failures such as invented file paths and unauthorised actions by rejecting any model output that does not match a strict JSON Schema. A path allowlist and command enumeration further limit what the agent can touch or execute, ensuring the system fails closed rather than proceeding on bad output. The tutorial walks through setting up a git worktree for isolation, defining the schema contract, and running pytest only after the proposed test command clears an allowlist. The author presents this as a reproducible, deliberately minimal recipe rather than a production benchmark, encouraging readers to copy and intentionally break the setup.