Mapping Goose to the Core Loop of AI Agents 🧬☕
From DNA of AI Agents to a Barista Coffee Agent
Every AI agent — from the simplest chatbot to a swarm of autonomous specialists — runs through the same core loop:
Perceive – sense the world.
Think/Plan – decide what to do.
Act – take action via tools.
Remember – store useful information.
Learn – adapt over time.
In my last post, I called this The DNA of AI Agents. It’s the blueprint that underlies everything from ChatGPT to complex multi-agent systems.
Goose in the DNA Loop
Recently, I ran Goose — an MCP-enabled agent framework by Block — and decided to map it back to this DNA.
Here’s what happens when you run Goose:
Perceive → Takes in your prompt and any relevant context from local files or prior steps.
Think/Plan → Outsources reasoning to an LLM, which decides what tool to run next.
Act → Calls MCP servers to interact with tools like Git, Slack, or custom extensions.
Memory → Stores session context locally; can be extended with vector databases.
Learn → Static for now — no self-improving strategies without developer input.
Seeing Goose mapped to the loop makes it easier to understand what it’s doing and what its limits are.
The Barista Agent
Of course, not everyone wants to think in abstract diagrams. So I gave Goose a job: make me a cappuccino.
Here’s the flow:
You order a cappuccino → the agent perceives the request.
It considers the recipe and tools it needs → think/plan.
It acts by preparing the drink → act/tools.
It remembers your preference for extra foam → memory.
Next time, it uses that memory to make your coffee just right → learn.
Whether it’s coffee or code, the same DNA applies — only the tools and skills change.
Why This Matters
Understanding this DNA loop and the five dials gives you a universal lens for looking at any agent framework: Goose, LangGraph, CrewAI, AutoGen, or the next thing that comes along.
It strips away the hype and shows you:
What’s actually happening.
Where the framework shines.
Where it needs work.
Once you see the DNA, you can build, compare, and design agents with much more clarity.