Zero to Agent in 30 Minutes: Build a Workflow Agent with John Berryman

When I first clicked on the O’Reilly Radar piece titled “Zero to Agent in 30 Minutes,” I expected another lofty, theory‑heavy overview of large language models. Instead, I got a down‑to‑earth, step‑by‑step demo from John Berryman, an AI consultant who spends his days building real‑world products for Arcturus Labs.

John starts by nailing down what he means by an “agent.” It’s essentially an LLM wrapped in two loops. The outer loop is the familiar chat you see in every chatbot – the back‑and‑forth with a user. The inner loop is where things get interesting: the model decides which tools to run, executes them, and then feeds the results back into the conversation. In his words, that inner loop is what separates a polite assistant from an autonomous agent that can actually get work done.

To make the idea concrete, John walks us through a tiny but fully functional review pipeline for job‑candidate applications. He doesn’t just talk about it; he builds it live, pausing now and then to point out why a particular choice matters. The result is a repeatable "outside‑in" recipe that any developer can follow, even if you’re only comfortable with vanilla Python.

The recipe can be summed up in five clear steps:

1. Lay the foundation first. Build the non‑AI parts – the UI, the data schema, the storage layer – before you think about the LLM.

2. Fake the AI. Drop in a static stub that returns canned responses. That way you can verify the rest of the system without waiting for a model to fire up.

3. Swap in a simple agent. Using Pydantic’s agent class, replace the stub with a real LLM that emits a structured payload (fields like update type, internal notes, and correspondence). The structure keeps everything predictable.

4. Give the agent a toolbox. A tiny suite – read, write, edit, and a shell – is enough. The model already knows the basics of bash and CLI tools, so you don’t need a massive library of custom functions.

5. Distill intelligence into a natural‑language skill. Write a concise prompt that captures the context, decision criteria, and step‑by‑step workflow. Sprinkle in a checklist so the agent can self‑validate before it hands you a final answer.

During the demo, John pauses to note that this isn’t a one‑off hack. He predicts a future where agents gradually replace a zoo of purpose‑built apps, consolidating functionality into a few cross‑tool companions that act on our behalf. It’s a bold claim, but watching the pipeline spin up in under thirty minutes makes it feel plausible.

If you want more of John’s take on the subject, mark your calendar for July 23. He’ll be presenting a session called “Escaping the Harness” at the AI Superstream conference – the venue isn’t spelled out, but the vibe promises a deep dive into freeing agents from overly restrictive scaffolding.

The article also teases the next episode of the O’Reilly Radar series, where Craig Hewitt, founder of Castos, will build a “complete social‑media agent” live using Hermes – the same architecture that powers OpenClaw. And for those hungry for a longer immersion, there’s a three‑hour AI Codecon workshop on open‑source AI tooling (registration details were linked in the original post).

All told, the piece feels less like a marketing fluff piece and more like a mini‑workshop you could actually replicate on your laptop tonight. If you’ve been curious about agents but felt intimidated by the jargon, John’s walk‑through gives you a clear, human‑friendly foothold.