I've been playing around with agentic AI for a side project recently, and I can absolutely see a world where the workflow is like this:

1. I have an iterative session with an AI to define a feature - the outcome of this is an AI generated brief
2. The AI goes off and implements the feature, unsupervised, the outcome of this is a pull request.
3. I review the pull request - what I'm looking for is a cursory look over the code to check that the patterns are holding, and brief manual verification that the feature is working.

This is not my current workflow - that step 2. I have to supervise - because the AI goes off into rabbit holes and burns all my tokens and creates some messes.

But this workflow seems tantalisingly in reach - whether that's all a fugazi is a different story.

I have limited time. I work a full time job, and I have a four-year old and two four-month old babies, and I need to exercise. So with my side project, of which I'm about 40 hours in, I could spend my limited time chipping away at building new features, or I could spend it:

• Optimising my prompts
• Adding lint rules and other guard rail test scripts
• Adding generation scripts

All in the hope that at the end of it, in some future world - creating this application will be a delightful breeze - I'll be able to informally have a chat session with an AI, creating a list of features I want to see, and then come back the next night to review and merge the code it has created.

This feels like yak-shaving. Perhaps this is another bit of software engineering wisdom that is disrupted by AI.

The AI maximalist might argue that - 'This isn't yak shaving - this is what the job is now, you're not building products, you're building a factory that builds products'.

This might be a case for a project at a high level of AI maturity - but for everything else there needs to be an appreciation that agentic AI workflows will not be a free lunch.

Questions? Comments? Criticisms? Get in the comments! 👇