
Vibe coding is perfect when you're still finding out if anyone wants your thing. Once real usage and real money show up, agents make software engineering matter more, not less.
If you make software with meaningful usage and money changing hands, you need software engineering now more than ever. Yes, even if you are using teams of agents to do it and you aren't sweating the actual lines of code much.
Building something new trying to see if it's going to work? Or something just for yourself or your team? Vibe code your ass off. Isolate the environment but don't worry too much about prompt techniques. Just talk to the agent until you get what you want. There are some minor tricks to it and security risks to manage. But mostly you're trying to get people to use what you built at all. It's a much bigger step beyond that to get someone to spend even a dollar on it.
If people are spending real time and money on your thing and you're past "explore" mode into "expand" mode, you'll want to work differently.
Break down the work into digestible pieces, sweat the acceptance criteria of each piece, and make sure that agents can verify those criteria before you spend human time doing it.
And you should spend human time verifying after the agents have signed off. Even with smaller changes the smartest agents still often make really odd choices that result in an off putting experience for most regular humans. Even when they still technically satisfy the acceptance criteria.
Digestible slices. Clear criteria. Automatic verification. A testing and regression foundation. Production observability and good feedback loops. "Gardening" workflows to keep things clear and coherent for agents and people over time.
Get these right and you can really scale up the amount of changes you can safely ship. At a much faster pace than the teams of a few years ago. If you don't, expect to gradually and then suddenly ruin what you already have if you try to ship the whole roadmap in weeks.
Despite all that we have the ability to do now, my experience of a lot of software I use is that it is getting worse, not better. I suspect this is because most software companies are applying agents badly and / or recklessly.
Make sure you're covered on good software engineering, product, and ux and be the one making yours better, faster than we ever dreamed before.
Agents can run multiple rounds of review and verification before you look. But ship to humans without looking yourself, and something subtle always feels off — because fully specifying a system that feels right to a human is hard.
AI agents get you to 90% of a working product faster than ever - but the real challenge is the other 90%
Persona-style agents are fun to name, but juggling Claude Code, Codex, ChatGPT and more calls for a single interface. Right now the answer might be to build it myself: a filterable Kanban board with a direct line to each agent.
Get More Like This
Follow along as I build and share what I learn
Found this helpful? Share it with your network!