Human Review Stays in the Loop

to be honest with you, I would say 100% of the time human review in the loop is absolutely necessary. with with with in terms of autonomous loop, when you go granular, then yes, autonomous loop is good, but you know, when you have a while kind of wild layer, you still need to consider the model which you are using is going to be probabilistic, it is not deterministic. So, how does a loop exactly work? Is the same. You you you run in multiple loop, like okay. hey, you execute it, and this is my verification. Now it will run one time and it is not passing, then you will share. Okay, this is the error. So next time it will have informed decision. Next time it will run, the verification layer will be run. The next time it will run, the third time it is able to make sure 99% of the accuracy is there. The fourth time it will try and run, it will be able to reach 100%. Now this is just auto automated verification. Like if the coding agent itself is doing the verification, will it be enough? No, definitely it's not enough. you still need to provide context in time to the coding agent what kind of test you need to write and what kind of verification you need to write, but you need to have an independent verification both autonomously as well as manual, especially when you are actually deployed to production, then it is always better to do evaluation both automatic fashion as well as manually. So that combination is very much required. How much automation depends on your company and your company's you know fault tolerance and all those things. so you can always decide if it is a solo engineer who is doing it. many of the cases like UX changes or UX changes might not require very strict process, but if it is a stripe related changes, then it might require a manual testing in place. So human review will definitely be required. once it goes beyond a single engineer in a team environment without human review, I think we will we will go into a mess. What I shown as wiki, for example, as an engineer, I might have my own flavor. What if I kind of try to dominate that wiki through my coding flavors indirectly? Like, hey, you don't do this, do this way, and automatically the coding wiki gets keep on updating, and I will be able to corrupt in a month's time and so on. So it's not not good maybe not good at this point. We don't know the future. The especially in AI engineering, the future is like evolving every day, even I am learning. probably when I started, just sharing one minute of what my process was when in 2023, when we are building AI product like suggested response and so on. it's like a small chat board with with rag and all we evaluate based on like we create a golden set in an Excel sheet, and we will evaluate. Okay, what is the output look like and is that output is matching to the golden set? what is the metric for that evaluation? And we will be manually verifying that. But now the evaluation shoot itself on the enterprise level has grown quite huge. you know, you can trace each and every behavior and things like that, and you can auto-evaluate, and you can even pinpoint like where exactly these autonomous run is failing, and what are all the bottleneck. There are many startups which are working towards that Langfuse, Opik, for example, and even Mastra have traces and observability. So on an enterprise scale, there are many many big players which are actually working towards that, and even then they themselves are not saying human in the loop is not needed. So that means it is still needed on every scale. done fast doesn't mean quality. So quality is something we need to check on that.