I’ve been doing lots of work on DAG Studio, and I find all my current tools lacking. I am using the Continue VSCode extension most of the time, which I like, but which also has some serious limitations especially with current Gemma 4 models.

This wasn’t a “seamless integration” or a “smooth journey.” It was crunchy. I’m talking long hours, late-night debugging sessions, and a constant loop of generating code, inferencing it, seeing it fail, and hammering it back into shape.

The Evolution: Moving past the AI-CLI

If you follow my work, you know I had a similar vibe going with my ai-cli project. That was a great start, but it was built on pure LangChain with a custom RabbitMQ setup and a lot of manual logic to keep things moving.

I was basically attempting to build the deepagents framework, but I was doing a much worse job of it.

By switching the backbone to deepagents and swapping the messaging to a Redis queue, the architecture shifted. My original vision at its core was an agent and queue of work. I had a lot of ideas about how to manage the context, prompting, tools, and more, but all of that is now taken care of by deepagents.

The “Crunchy” Phase

The transition from “it mostly works” to a reference implementation was half the battle. I spent the first part of my week prototyping and getting things working in a rabbit hole of tool-streaming and context management. I spent the second half of my week cleaning up the prototype and polishing everything to a reasonable level.

I’m using Gemma 4 (via llama-server) as the brain. The power of the Gemma 4 family—especially being able to route between the 4B for fast processing, 26B for triage and the 31B for heavy lifting—is what makes this actually viable.

I built a modular Agent Factory with a SwappableLLM layer using a ServerManager to manage llama-server for local model swapping and LlamaServerAdapter to handle Gemma4 and my own requirements. It was a slog, but the result is a system that can actually perceive state and handle its own persistence, and is a solid starting point for almost any project.

What is Agentic Core?

Basically, it’s a blueprint for anyone who wants to build an autonomous environment. It’s not just a wrapper; it’s a full operational stack:

• Dynamic Routing: Swapping models based on the complexity of the task.
• Persistence: A dedicated /memories/ directory so the agent doesn’t have goldfish memory.
• Decoupled UI: It currently talks to me via Discord, but since it’s all Redis-backed, you can plug in whatever interface you want.
• Worker Process: Watch the worker reason / form tool calls / process tasks by just running the worker in a terminal.

You can check out the full implementation here: Agentic Core on GitHub.

Danger Will Robinson

Just a reminder: Giving an agent the ability to schedule its own tasks and execute tools is powerful, but it’s also a great way to accidentally create an infinite loop that spams your API credits or your Discord server. Keep it sandboxed until you trust your prompts.

Support the Build

Building these things is a massive time sink, and I’m doing this as a personal lab for my other projects. If you find the patterns in Agentic Core useful for your own agency needs, feel free to buy me a coffee on Ko-fi. It keeps the caffeine flowing and the debugging sessions going.

Now go build something amazing.