SpeedRun Ep 06 Durable Agent Writing -Tales from the Sky Lounge

All right. So the next thing I started looking at is what do these things look like? Not everybody's got Claude Desktop. It's really cool, rough and ready thing that you can connect a lot of MCP tools now that you have remote desktop. It's got web search, it's got deep thinking, it's got planning, it's got memory, you can set up projects. Very cool. But as a CTO of software companies, there's times when that is not durable enough. How can we take what we've done on the desktop and then make it a little bit more permanent? And is there a really slick way that we could take some of these chats and do something with it? And I thought, man, okay, this is definitely got to feel like close to the edge of what Claude can do. I was going to give it some easy stuff and see what happens. So I was pleasantly surprised. This is not the end of the rainbow here. How do I create a durable agent based on Claude with MCP tools that can be run on a schedule or as a result of an event?

So I went off and thought for exactly 22 seconds and it said, "Here's what you do. Consider some of these frameworks. Get you a database." Okay, not shabby. And then handle some stuff and then think about MCP, go get you an SDK with Python or TypeScript and then configure all this stuff. Then it starts writing code. So here's a Python script and it goes through and shows how you it would look in Celery. Okay, not the end of it, not knocking my socks off yet, but here's some options, right? Docker, Kubernetes. Okay, not too bad. Cloud Functions, AWS specific. Very cool. DynamoDB. Okay, Firestore. A little GCP action there. Traditional server and job queue. All right, these are all things that I'm used to seeing as a CTO of software companies. So here's a little bit more thinking code. Okay, so he's showing me what they're thinking. Error handling. Yeah. Okay. Durable security and off definitely. And I said, okay, is Pydantic a viable agent framework? And it said, not complete on its own, but you're going to need some other stuff. And it's excellent. By the way, it always tells you you're excellent. And that's awesome and you're a good person and what a great thought. I don't know, it doesn't hurt the ego I guess. So little bit education there and then we go down and then I said, can I, this is actually an interesting question. Can I rapidly prototype an MCP agent in Claude Desktop and then convert it using the configuration and the prompt to a production agent? And it said actually yes and it's brilliant. Actually, it is brilliant. Claude Desktop makes this perfect. Here's the flow. Configure your tools, design the system, and then there's a certain breakdown to where all this stuff works. And then extract the production configuration, and then if you start squinting at it, you can see that it's starting to back into LangGraph and then Pydantic and what they're going to want, right? These are the things they're going to have to go into it. And then convert to production framework. This thing just shifts into code mode like a beast and then it starts writing code and then it said blah blah blah and it says hey you want me to show you how to extract stuff from the prototyping and I go why yes I would.

And then so this is kind of a thing I discovered is it's going to walk you through and checkpoint and then ask you a question then you go yeah please continue or do something different and then you know it's surprising how many times you just go yes yes yes yes yes and it takes you to a really cool place. I think this is a place where you got to use your judgment and figure out where it's going, if that's really a place you want to go with it. And it said, hey, I'm going to think about it for a second. And then it said, here's some extraction. And then here's an example of like a desktop session. And then here's what we would extract. Capture the tool use. That's your MCP configuration. Actually, which tools did you hit and which methods? Convert to system prompts, structured system prompt in Python. Decision trees. Oh, this is kind of cool. It's going to go watch you. Oh, that's a very respectable regex there, if I don't say so myself. Codify handling patterns. This is all very cool. Lots of little code snippets. Yeah, pretty cool. Okay. Validation. And then it says, "This approach let you maintain exact behavior. Would you like me to show you how to implement anything?" I said, generate Python code to take a Claude chat output and create durable agent using Pydantic and LangChain or LangGraph from LangChain. And it says 4 seconds. Here you go. There's the code. It's in Python. It's pound bang user bin blah blah blah. Look at all this code. It wrote a freaking tool for me.

This code provides a complete system. Here's how you would do it. And example convert agent. And then you basically paste in a chat and then you convert it. And then oh by the way, just in case you're power agent, throw everything in a folder and then just point our newly created piece of software at that. Convert chat to agent blah blah blah and then name it agentname agent.py and then you could go throw it up on your AWS or Azure infrastructure, whatever you like. How cool is that?

So two important points. I love code, so as a coder, I can respect that and I can talk to it and I love it. I know how it thinks. That's awesome. You got to be careful. You got to look at the code and you're going to want to run it and make sure to test it in small ways, but stuff I've seen has been pretty dang close with very few little nits to go through. And then the other thing is you can go from Claude Desktop, have your users go through and get good at a particular workflow automation, maybe put in a project and then hit next next next, handle a few errors and then convert it into Python code which you can deploy in production environments. That is how we're going to get to a million billion agents. More, 10, 100x more agents than there are apps in the app store according to Jeff Haney on one of my last podcasts. Totally see it. Totally get it. Agents for everybody. The cost of agents is going to go to zero. If you're in business to try to make agents for people, that's not going to be a business in about a year once people get a hold of this technology and self-serve. Anyway, that's my thoughts.