I’m starting to understand that it is actually a game of volume only. It doesn’t work if you are trying to be too clever with it. I was trying to automate making scripts using agents. Right? So I made one skill. It would fuck up. So I made another complementary skill. It would still fuck up. I made another complementary skill. It, like, cascaded into, like, four, five skills, and it would still miss out on something. Okay. But then I made one skill only which has appropriate context and then said… you will have this checklist. If anything is noncompliant with any item from checklist, you make change, and then you will have to run it again and check if it is compliant with everything else. It had to do, like, twenty, thirty rounds of iteration, only then it was able to comply with everything. And that was such a fabulous script. And it wouldn’t happen with a Hermes agent because it it has propensity to stop too early. It was possible to do on deepseek webapp because it thinks for a long time.
So that’s what I understood. that even in business, you can’t be trying to be too clever. You just have to do the volume. You just have to do the gruntwork.
And in case of agents it generally means that having a checklist and when it generates an output checking it against that checklist and even if there is one non-compliant thing and you make some little changes you have to run the whole thing through it again really strictly and it might find another issue with it so you just keep on going and going and going until all the pieces fall into the place. It’s almost like Sudoku where at start you are adding number randomly and then it all converges towards a solid base but you cannot stop unless until it solved.
I wonder how many AI related things I gave up because there was just not enough volume performed and I tried to be clever with it with use of multiple skills and constraints when one shall not have many constraints but just iterations until all of them are satisfied so that it can give you something really really solid. So I just first need to figure out which AI tools I can even use. Hermes is not built for this. It is- it is going to fuck me over by just doing like three four things out of 11. Forget the iterations. So what should I even use? I have to think about it now
The core requirement is: long thinking + willingness to iterate many rounds without stopping early.
This is almost as if I rediscovered gradient descent, where basically you have one function right where you want to minimize loss so it will keep trying to travel to its neighbors to try to get to the minima and in this case what is happening is I have a set of constraints and I want all of them to be satisfied with so it’s like we have to travel through all this space so that we end up at a point where all the constraints are satisfied so of course it’s going to take time, this is essentially gradient descent where assuming you have 16 constraints, you can either be in compliance or non-compliance, so your score can be minus 16 to plus 16. Assuming minus 16 means you are in compliance with everything. And if the data set is complicated enough descent can take months. so if your constraints are complicated enough you are going to also need a lot of compute, iterations and time to get there you cannot be impatient there is no way you can one shot this
This is especially true when you are trying to write some article Or trying to get something visual made using code like Manim animations.
My fix for this whole situation is to have two agents one is an auditor whose context will reset after each audit. It will have a list of lets say 15 things and it will audit from granular to the macro level and the interconnection of it all. And then you have the writer which will retain in the context. The auditor will have the incentive to be as ruthless as possible with its judgment and the writer will retain the context and see how the situation is evolving so it can may be improved with its response
What are the techniques I can steal from gradient descent that can help me navigate it. It is a really really well researched field right. So what can I steal from there to help me since the problems are so similar.