Developer shares notes on wiring AI-SDLC workflows with agents versus scripts

On November 5, 2026, a post titled "AI-SDLC Practice: Should the Workflow Be Orchestrated by an Agent or by Scripts" was published on Juejin. The author reviewed two extreme approaches to structuring an AI-SDLC workflow: handing the orchestration to an Agent, or chaining it with scripts. After trying both, the author settled on a middle path: letting an Agent operate the CLI rather than fully scripting the pipeline or leaving the entire sequence to Agent discretion.

The write-up goes on to describe how the author divides the pipeline into Workflow, Phase, and State, how skills are bound to each stage, and where human confirmation is inserted. Six human checkpoints are defined, and hook events are designed to trigger at key moments in the flow. Together these mechanisms aim to keep an Agent driving the CLI while still giving a person explicit points to review and approve before work proceeds.

Only this one article has been found so far. The piece is framed as a personal retrospective of the author's own setup, and no separate independent coverage of the workflow, its adoption, or its results has surfaced to date. Details such as which AI coding tools were used, how the CLI was driven, and whether the approach was measured against a script-only baseline are left to the original post.