I recently got to participate in a roundtable discussion. The format was interesting — people gather and talk through one big topic. No slides, no presenters. Everyone's equal, just a conversation.
It was my first time in this kind of setting, and I ended up seeing a lot of things from a completely different angle. Hard to overstate how useful it turned out to be.
The topics were pretty heated:
- How to bring agent-driven development into real workflows
- How to make it predictable and quality-controlled
- How any of this survives in big tech, where every process multiplies across dozens of internal constraints
It's clear the industry is heading in this direction. The "ask ChatGPT → copy the answer → paste" workflow doesn't really work anymore. What's starting is a more complex story involving agents, processes, and control.
What we roughly agreed on
Pet projects don't scale. Personal experiments with AI skills almost never translate to the department level. There's too little control, which means too little predictability.
Internal "vibe coding" clubs actually help. Workshops and AI experiments inside companies turned out to be genuinely useful. Even failed experiments give people real hands-on experience with AI tooling.
The new bottleneck is coordination, not production. We can now produce far more solutions, but shipping features is still the bottleneck. When a single release has dozens of features in the pipeline, the question of control gets sharper, not easier.
You don't have to be first. It's not always profitable to be the earliest adopter. Often it's cheaper and calmer to be somewhere in the top ten — learning from others' mistakes, with clearer economics.
The quote that stuck
One thing someone said during the discussion really stayed with me:
We all switched from shovels to tractors, but the field is still the same size. Now we'll just bump into each other more often.
We can build more features. Coordinating all of it — that's still the hard part.