Siraj Raval

The future of agent interfaces is not walls of text. One rising solution to this is OpenUI®. Another is json-render. Both are good tbh. One thing that I wanted to look at was performance. OpenUI claims a 67% reduction in tokens. When I re-ran the benchmark I didn't get that. There is a savings in tokens. But to me it does not seem worth the push towards a commercialized product, a smaller ecosystem, and leaving behind very well understood formats like JSON that agents have deep XP with. If you are building a product that uses structured components in AI responses, what do you use? ps OpenUI® I'll retract this if you convince me. 👀 Just not convinced your benchmark is remotely accurate.

6

My hot take of the week: in the age of coding agents, the best way to review the 101829038120 PRs your agent is making is through… video. Zach Zhong and I ship constantly, and opening a PR is basically free now. Writing code is cheap. can’t say the same about reviewing it. You can’t feel a UI change by reading a diff. And too often, we found ourselves acting like the robots… manually running e2e checks to make sure one PR didn’t break something ten screens away. So we built an internal tool where every PR comes with an auto-generated video: a visual walkthrough of the full pipeline, paired with captions. The recording gives us a visual artifact to collaborate, catch issues, and ship faster. Funny bonus: the generated video can also become a polished demo I can post or send to clients. Yay auto-distribution For people building with coding agents: how are you reviewing all the code they generate? (btw as seen from the vid below we shipped environment variables configuration for workflows built on bubble lab!)

59

10 Comments

Several people know about this problem but nobody talks about how to solve it. I was asked one time to help fix this issue at a company. The fix lies in breaking the work down to a few critical steps. First, never just hand off work in its entirety and expect you'll get anything you can use. Regardless of how good or bad the team is, always ask, always, the offshore team to pseudo code their approach. They should explain which files they'll touch, which systems they'll touch, integrations, which areas or layers of the application, etc. I call this the first pass. Why is this important? This tells me they're thinking about the problem right and if there are any gaps in their understanding, or any pieces they're missing. After that, they'll do a second pass. That is stubbing out which functions or code will need to be added or modified, maybe even making the functions and service calls work but with mock data. This is the "prove it to me" moment. This second pass also helps me as a dev manager gauge how fast or slow the offshore team is working, giving me indications too if they're stuck somewhere. Only after this second pass, do I actually turn them loose to go execute and "deliver" the solution. Only now, because I did some diligence on my managerial end, I've got greater confidence that what I'll get will be something on par with what I'm expecting and can use. This partnership and collaborative arrangement also fosters trust and over time the offshore team will learn how your organization operates, critical thinking, and that nothing should ever be done in a vacuum or cowboy way. For more helpful tips like this, follow me. I'm paying it forward in 2026 and will be sharing gems most people aren't talk about with a focus on the solutions, not just problems.

1