The Agentic Engineering Workflow
A practical, eight-part guide to moving from ad-hoc AI prompting to a disciplined workflow you can rely on in production. Each part is self-contained: read it, copy the examples, and set up that piece of your own workflow.
The thread running through all eight: your real output is no longer code — it's the system that produces code. The model is one small part of that system. Everything you build around it (rules, context, tests, review, observability) is what determines whether the output is trustworthy.
How the workflow works
The series
- Set up the rule file — give the agent the project knowledge a new teammate would need.
- Engineer the context — control what the agent sees, and when.
- Build verification — tests and evals as the contract with the AI.
- Run the work — conductor vs orchestrator, and where agents fit in your day.
- Review and ship — catch the failures that "look right."
- Control cost — total cost of ownership and model routing.
- Ship production agents — from a prototype script to a product with a substrate.
- Make it a team standard — version the harness, gate on evals, hire for judgment.
How to use this
- Solo developer? Parts 1–6 are enough to transform your daily workflow. Start with Part 1.
- Building an AI product? Add Part 7.
- Leading a team? Parts 1–8, with extra weight on 3, 5, and 8.
Source: The New SDLC With Vibe Coding (Google)