ADK Quick Review

 Got into a new job and i had to switch cloud vendors, so part of my learning curve is to get involve in new SDK's like the ADK from google, I havent paid much attention to that one as i thought it was Gemini only but it seems we can run locall LLM's with it and take advantage of their features.
So, I have decieded to run a quick demo to see how fast we can implement a quick agent with tools. On this example i will let the Agent talk to Docker API on the system.

 

Now, let try let the LLM look into logs, find errors and provide suggestions:

 

 

This was a quick night after work so i didnt have much time to set it up but i am quite impress with the tracing/sessions/states/artifacts this is quite the framework to work with Agents it does makes things much easier at 1st glance.
Ofcourse on this sample the LLM its a bit confuse and tries to talk me into using the CLI but that's just because i didnt set a proper prompt i was testing how fast we can get on track.
I switch to an MCP server as they are too convinient to move around Agents and worked right away.
I'll try to create some better workflows and test the A2A framework in another time.

From Overwhelmed to Efficient: Qwen 3 and the Future of Ticket System Management

 Qwen3 recently came out and It seems the 1st thinking model with decent results. I was not too impress with DeekSeek results running locally, so i have decided to give it a try.

For this model i have decided to mimik a ticket system for Lv1 Support and create a simple agent with Qwen3 that will be integrated with the ticket system to manage its tickets through actions.

I am going to try 2 aproaches here:

• Support Engineer uses an agent to work on his queue.
• Ticket owner works directly with the Agent. 

 So before I start playing with Qwen3, I need to create a basic ticket system that will allow the basic CRUD operations and decided to use Flask & sqlite to quickly prototype it.

And this is how it looks basic but good enough:

 

Now, that we have the ticket system ready, i can start integrating the LLM to perform actions.

So a visual representation of this would look something like this:

 

This is a very basic implementation but the goal here is to see how good the open source models that run locally have become.

After drafting some code I got a basic agent to handle the basic CRUD (Create/Read/Update/Delete) Operations on the ticket system.

Here is a quick test:

 It was very interesting to see the LLM been able to answer the questions and interact with the ticket system from a basic agent.

This shows that there could be some uses cases for engineers that work with a ticket system, they can have an agent working with them on the tickets they have assign to speed the resolution process.

 Now, if you look at the industry for AI its quite focus on chatbots like the previous example but from my personal view, there is more power on removing the user interaction and integrate the AI Agent directly into the pipeline, something like this:

In this way the ticket creator will interact directly with the Agent.

So after a few code changes I have intergrated the LLM into the ticketing system.

In this way the ticket user will interact directly with the Agent and the agent will help with the current issue and any new issues that show up during the troubleshooting process as a regular ticket interaction.

For a demo this is pretty nice but currently all the answers from the agent come from its own knowledge. So in a real corporate/enterprise scenario We cant use this, We would need the LLM to know about propretary data, like company processes, policies or an internal knowledge base.

So I guess the next step for me is to find something similar to that in the public internet and integrate it but maybe for a part 2.

I might need to stop here and change the LLM as anthropic recently publish a paper invalidating thinking models:

 Reasoning models don't always say what they think 

So if the reasoning is not accurate we might be spending time/tokens/money for no reason.