Another insane use case with Claude Code | Self-Improving AI Agents

Key Takeaways

Hey everyone. So yesterday I posted something about sub agents and how you can combine that with flash commands to get more reliable results from your cloud code sub aents and I actually wanted to explore this idea a little bit further. What more advanced agentic architectures we can explore. One idea that I use a lot is this idea of evaluator optimizer. The idea is that we have an evaluator which assesses how good the content that's generated by our agent or your agentic workflow. In this case, we have a really simple example of an agent that creates a post given a URL. So, this would be something like for X or any other social media platform. So, this creates a post draft. The post evaluator assesses it based in some criteria rubric that I've provided the agent. Then, it provides some feedback. If the post generator generates something of high quality, then we are going to approve that post. And finally, we're going to get an output. So, how to orchestrate something like this inside of cloud code? That's what I want to show in this video. So here I'm on cloud code and I've already set up my post evaluator agent and post generator agent. The post generator agent basically just creates this task to generate a spicy engaging and technical post for the X platform and it needs to keep the length to 200 to 300 words and then I have the post evaluator agent and this one is tasked to assess the output of the post generator agent given some criteria. it needs to produce a score between zero and 100 for each one of these dimensions and then provide a final score. Once we can produce that final score, then we can sort of assess whether the post that was generated by the post generator can be approved. I did not put that logic here because what I found with any evaluator agent that I've created, it tries to take shortcuts and introduces biases. For example, it was just output a really high score just to meet that criteria. So I wanted to leave that logic out of the agent and put that somewhere else. So where would I put that? And this is where the idea of slash commands comes into play. So I have the logic baked here. But I also have additional highlevel instructions for how these agents should interact with each other. In this case, because we want to generate the post and also use this evaluator to provide feedback. All of that instruction is here. And notice that I have this loop here which is iterate over a step one to three until the post evaluator agent grades the post with a score of 90 or higher. So that logic that I was explaining is something that I baked into this flash command. If we get an approved post, we'll save that in a folder. And that's pretty much it. So this is building on that idea that I posted about yesterday of using sub agents and orchestrating the way the sub agents communicate and complete tasks. In this case, I am interested in creating this loop. So I'm testing the idea. This is the first iteration of it. So it's not perfect, but essentially what I want is I want to be able to create more complex interactions between the sub agents. And so I'm not going to leave that orchestration to cloud code to decide because this is a really custom agentic workflow that I'm interested in developing just to show you how that works. Now I was just use the slash command post generator and then provided a URL. The URL is now provided as an argument and the two agents can then start to perform the task. So the post generator would generate the post. The evaluator agent will evaluate the post. Then the evaluator agent sends that feedback to the post generator and then we do the iteration order loop that ensures that on every iteration hopefully there is an improvement on the post. Let's look at a concrete example of how these sub aents interact with each other and whether this actually works. You can see here that the post generator generated a post and then the post evaluator evaluated it. It gave it a 77.5 score which doesn't meet the 90 threshold. So then it's going to send that back to the post generator as feedback and then it's going to try to improve. You can see here improve post based on feedback. One thing I want you to pay attention here is the number of tokens. Because this use case is sending back and forth information, you could expect that the amount of tokens used is going to be really high. So be very careful when you're playing around with this particular architecture. There's a lot of little optimizations that I can implement between these sub agents. But this is the first iteration. I definitely have a lot to improve. You can see that the post generator, the post evaluator, post generator, post evaluator again, and it scores very close to 90. And then it says okay I need to go again and iterate over this. So it generates the post again. It does the evaluation again. You can see that now it's 83. After a few iterations something really interesting happened. Basically the agent decide that it was consistently getting a score in the 80s. So it decided to just finalize and then save the result. Finally the result was saved. Now this is not ideal because we still want to meet that 90 threshold. But what I found out is that it's really hard to actually get over 90. And so in a way you can argue that this is okay. But if you really want to be more strict, you should definitely instruct it here. It needs to be over 90. If it's not over 90 after like 10 iterations, then close the loop or something like that. You can put more conditions if you want. That's the power of having detailed instructions. So after that, it's just save the post and we can take a look at the post here. These are just the scores, but let me just show you the output here. And this post is related to the famous Antrophic blog post on AI agents. So definitely spicy. Antropic just dropped the nuclear take. Every dev needs to hear the dirty truth. your rap plus function calling not an agent that lchin monstrosity workflow pretending to be smart real agents equal systems that control their own execution flow but these are just the design patterns and then there's some reality check then hot take your intelligent system probably needs human in the loop not another abstraction layer that's funny who's ready to delete their overengineer agent stack and start fresh so a lot of really spicy takes here and I really like the output so far but definitely can be improved as I said this is just a first it situation showing you that you can do more with slash commands and sub agents inside of cloud code I hope you enjoyed this content. Let me know if you're interested in more videos like this below. And I'll be doing a full breakdown of all the agents and some other cool ideas for my Pro Academy members. So, I'll see you all soon.