15062026 CHAPEL VALERIO BARBERA

So, let me open all the stuff I need. And I need to reopen later. So, yeah, the title of the talk is quite explicit because it's agent building agents and it try to explore an intersection of two very important trend that the software development industry is facing recently. And it's the two way the AI is affecting um the the the software developer work. And this AI is changing how we build and the AI is changing what we build.

The how we build is something that maybe someone of you experience in working coding industry. Uh a lot changed in the in the last year basically because in late 2025, in December 2025, where the the most capable models were released by company like Anthropic and OpenAI, these tools, the coding agent tools, started to be to become uh really useful tool. They started really working in write reliable code. That's was the real shift because before uh the models released in that period, uh these coding agent were just I don't want to say a toy, uh but they are useful only to work on very limited task, very simple uh code implementation, maybe single functions. And with Opus 4.

5, basically and GPT uh 5, this kind of tools became very powerful. And now, as a software developers, and I I am an as real example of this of this trend because I am embracing this new way of work uh at full speed. I'm now a a new collaborators that recently uh joined my my company, and we two, me and this collaborator, are working with roughly two, three coding agent in parallel the entire day. So, we are now uh basically we can compare ourselves with a team of four, five people. And this is something that really scare most of developers I talking to.

Uh because it's a very profound change in the way we we work, in the way we work at for, uh you know, the last 20, 20 years. And this is also very controversial. Recently, I saw on YouTube uh lot of videos of very important personalities speaking at about this change in the software development industry uh that were booed basically uh from their audience, mostly very young uh, audience because young people are the most uh, scared the most about this change in this way, uh, developers work. And the other change that it's happening in parallel is what we build is changing because in late 2024, in the beginning of 2025, as software developers, we are the first that are exposed to these kind of technologies and changes. And as software developers, we're the first that understood that this kind of technologies can not only be used to empower our everyday work, but we also can use this technology to build a completely new class of software product and, you know, integrations because AI agents are not only uh, coding agent or the product that these that the company building this technology are making available to us, but age agentic software is something that we can now build inside our software products, our software applications, to provide a completely new user experience to our customers.

So, the question was, what if an agent can build an agent? Because at the end of the day, it's the same skill. When we start learning how these kind of technologies work, using these these tools, the coding agent, the uh, you know, the AI interfaces that we have available, the chatbots. We are just learning how this technology works, and the next step for software developers is how to use the technology to leverage the same feature, the same capabilities to start building something new in our software products. So, I prepared a live coding session that I want to show you in real example how this is possible.

And we use two tools to do this demo. On one side, we have the the coding agent, so the agent responsible to write code and to build this new system, this software. I use Cloud Code because, you know, it's my preferred tool, but there are you know, plenty of them with different features, capabilities, open source, closed source, so you have a lot of choices. And on the other side, we need a framework, a coding framework that help us to build these entities with help of the coding agent I'm going to use. So, in this case, I use the Neuron AI framework.

It's the most important agentic framework in the PHP ecosystem. It's kind like something LangChain from the Python ecosystem. Maybe some one of you already heard about LangChain. It's most used agentic framework in the world, but it's on the Python programming language. So, Neuron AI was built basically on PHP vertical.

So, PHP developers can integrate agentic systems, agentic features into their application thanks to this to this framework. It allow it offers a lot of abstractions. Agent entities, tools, toolkits, memory, all the all the components, all the features that are needed to build complete agentic system. So, I'm Valerio. I'm CTO of this product company, Inspector.

dev, that it's an agentic application monitoring system. Based in Naples, but we have no customers in Italy, 100 of customers in more than 40 countries. And, you know, jump into the code. So, Okay, this is the make it bigger. This is the project that now I have open it in my in my code environment.

Okay, this project was built entirely by Cloud Code. I have registered a video of this coding the first coding session because, you know, at the live events, may connection, there are a couple a lot of things that can go wrong. So, I registered this short video just to just allow you to see this process ongoing, and then we can jump into the code and start working on the project with live coding and Cloud Code uh really. So, just a bit of uh information in advance. So, this is the documentation of the Neuron AI framework.

And this framework offer the ability to um make coding agent able to work effectively with the framework itself using uh skills. Maybe using this kind of uh agentic tools, you are really heard about skills, you're able to uh create the all the instructions that help uh the agent to work with your stuff, basically. It can be also legal tasks or you know, in this case, are just coding uh coding abilities. And the framework ships with this bunch of uh uh agentic skills that you can install in your project to help the coding agent write reliable code using this framework. So, in this video, I just started uh it's basically copy and past the commands from the Neuron AI documentation.

I installed the skills into the project. The Neuron uh test engineer, Neuron tool creator. Okay, so skills are installed. I open Cloud Code. And in Cloud Code, I past this prompt that it's basically the prompt I prepared to instruct the coding agent to build the the project I want to build.

And in this example, it's basically a business intelligent agent. So, I want to build an agentic entity that is able to talk with your database. Maybe you have an e-commerce application or whatever, a database is behind any kind of software or application. So, I just tell the agent to build this kind of agent that is able to talk with a database. So, in this case, a SQLite database with I give the agent a bunch of instruction for example, to use Here it is, to use Anthropic as AI provider because on Neuron Framework you can connect basically all kind of AI provider to power your agent with their models.

So, it's enough to copy and past this prompt into the coding agent. Now, here you can see I switched to auto mode. So, in this way the agent doesn't ask to me permission to read file, write file, just give you about the pool So, the video just I don't interact at all with this coding session powered by Cloud Code. So, I move the video forward because now it's all about Cloud Code ingesting and writing code. So, it's now writing the SQLite schema tool that it's basically what the agent need to first understand the schema of the database that it should read from.

Moving a little forward, now it's creating the business intelligence agent, so the agent entity itself. Connecting all the tools. Okay, it's doing just some check. Move a little bit forward. Okay, now fixing something.

The coding agent just treat continue to treat the code. The syntax if all PHP file are write correctly. And now here is trying to test the it's it's try it's calling the agent to see if it works. And at the end of the day everything works end-to-end. Here is the summary and the task was done.

So, the result of this coding session is exactly this project that I have here in my code environment. So, I can now open my terminal. Go into the project directory. I called it neuron BI instead of neuron AI, like business intelligent agent, because it's now able to talk with the database. So, I can just ask uh agent what are the top products?

Sorry? Yeah. Okay. Thank you. It's Monday morning.

What are the top products per month or the top product the top maybe three products just to query strictly. per month. So, we just launch these uh these question to this uh PHP agent. We'll see what it answer. Okay, I'll tell you the first two call is the analyze the database schema that it's a tool that we are going to see everything in the code.

Then, it it run a select which it talk to the database to extract some data and based on the data it gathered it's able basically to give us all the statistics and everything we want to extract from the database and everything it's in natural language. So, jumping into the code, we see here in the tools directory that we have a bunch of tools that we saw the agent use to perform this task. It's the SQL schema tool where the agent is able to understand the the database schema in order to be able to create the right select statement to talk to the database and extract data from. Now, the problem of most you know, a genetic system when you start building agent yourself, you face with a lot of limits that these models has by its nature. They are stateless.

So, you don't There's no way these models are able to remember the ongoing conversation you are taking with with it. So, if I ask the agent What are we talking talking about? Yeah, but they are smart enough. Okay. It should answer that it it has no idea.

So, yeah. Okay, we haven't started talking about anything. Okay. So, now I open Claude code and I ask Claude code uh file chat history to the agent. It uh storing uh chat sorry, messages messages in the history directory.

So, now we leave Claude code do its job and in the meantime Okay, I switch to auto mode here, too. So, I have no uh no need to constantly Okay, 2 minutes. We are we are done. So, what we are doing now is asking Claude code to basically add memory to this to this agent. Because the framework, the neuron AI framework has this uh uh kind of components that you can add to your agent that it's basically store the ongoing conversation into a single component, and you can use uh different kind of storage system connected to the agent.

In our example, we are going to use a simple file, but you can connect uh the memory of the agent with SQL databases or maybe other uh framework-specific integrations. Uh let me see where it's going to handle state. So, the agent should put Okay, it's going to write some code. Okay, it's adding some code to the agent entity to make it work. Here's fine finally found the file chart history component.

Uh it's mostly because we already installed the skills of the agent at the beginning of the project. So, the coding agent knows basically everything about the framework or components, how to use them. It's quite done. We'd see, and then we are able to uh talk again. Uh it works.

Let me create a test. Okay, but I think we don't need tests. Okay, it's done. Exit. Clear.

And now I first ask the agent the same question we sent before. So, what are the topics? Just to extract some information from the database. It start again to calling tools, running a query against the database, and extract all the information. But now, if I ask "What are we talking about?"

The agent now has memory of what we are talking with it. So, thank you. That's it.

Thank you, Valerio.

I'm sorry, but uh

That's okay.

we don't have so much time. Any question?

Any question to Valerio?