The ideal agentic AI software engineering workflow
Humans are done writing code. We already wrote a ton of it, and there’s more than enough now for AI models to be trained on it sufficiently to write the rest.
Sadly, many current workflows that involve AI writing code assume humans would still want to be involved in that part of the process. We have copilots in our editors, but that is just a way of introducing new inefficiencies, where humans have to spend cognitive power consolidating human-written and AI-written code — and generally have to think about the code-writing process at all. We ask LLM services to code, conversing with them until it’s just right, again mixing up mental processes and keeping the inefficiency of having a slow-thinking human involved.
Human software engineers should remove themselves from that process and instead spend all their time expressing needs and assessing results. Luckily, we already have every system in place for doing that — we already do it daily — and the ideal agentic AI software engineering workflow should leverage that while removing human code-writing entirely.
Using the popular software project planning tool Linear and the ubiquitous GitHub as examples, this is what it should look like:
