Building solutions with AI changed how I approach projects.
Hey Product Hunt 👋
I'm a lead software developer based in Switzerland, and lately a big part of my work has been defining systems and setups in order to harness the power of AI in software development and solutions building - and where it just creates a different kind of work.
Generating code isn't really the hard part anymore.
The interesting problems are everything around it: giving AI the right context, making good architectural decisions upfront, keeping larger builds consistent and knowing what should - and shouldn't - be handed over to an agent.
I've been testing different approaches while building real products, and I'll be sharing some of what works, what doesn't and what we're learning along the way.
Curious to hear from other developers here:
What's the biggest thing you've changed about the way you build since coding agents became genuinely useful?
Replies
The biggest change for me is that a task is no longer complete when the diff looks plausible; it is complete when the agent returns an evidence bundle. I ask for the exact files changed, assumptions made, tests run, test output, unresolved risks, and any side effects such as migrations or generated artifacts. For larger work, I also split planning from execution: one pass maps the existing architecture and proposes invariants, then a fresh pass implements against that pinned plan. The useful human review question becomes less ‘does this code look reasonable?’ and more ‘does the evidence prove every acceptance criterion without violating an invariant?’ That has made small agents easier to trust and exposed when a task was too ambiguous to delegate in the first place.