In the first post I showed a SUEWS-agent session end to end: I typed a plain-English request, and the agent built a configuration, checked it, ran the model, drew a figure, and left me to judge whether the result made sense.
So what turns a chatbot into something that can do all that?
If you have used ChatGPT, you know the chatbot pattern: you ask, it answers, and that is the end of it.
A chatbot is, in effect, a brain on its own.
It can think and talk, but it cannot reach out and touch anything, and it keeps no memory of its own beyond the conversation in front of it.
An AI agent is that same brain given a pen and a piece of paper.
The pen is the easy way to picture it: a way to act on something outside the conversation.
In practice, that pen is an engineered harness: a controlled way to act through programs, files, models, searches, checks, and the wider working environment.
The paper is memory, somewhere it can read from and write to that outlasts any single message.
Put the three together and you get a loop: the brain reads its paper to recall what it is doing, thinks, acts through the harness, writes down what it learned, and goes round again.
That loop, a brain with a harness and memory running until the job is done, is what we mean by an agent.
You saw all three in the first post.
The brain was the general-purpose AI model I was working with; the harness was its host agent environment with the SUEWS-agent loaded, the SUEWS-specific ways to create a configuration, validate it, and run the model; the paper was the configuration file and the run’s outputs, which it kept reading back to decide what to do next.
Figure: a chatbot is a brain that answers; an AI agent is a brain running in a loop with a pen-like harness and memory. Layout, arrows and labels were typeset locally with TikZ; icon assets were generated with OpenAI image generation from the prompt below.
Prompt used for the generated icons
Use case: scientific-educational
Asset type: icon sheet for a SUEWS-agent explanatory diagram
Primary request: Create a clean polished icon sheet with exactly four separate icons, no words and no letters: 1) abstract neural brain/network icon, 2) pen-like engineering harness icon, 3) document/memory page icon, 4) circular loop arrow icon.
Style: restrained scientific editorial, vector-like but with subtle hand-polished depth, clean line art, rounded geometry, suitable for an academic climate-model blog post. Use SUEWS-like colors: deep navy #123B6D, teal #006C73, muted blue-grey #8FAAB5, orange #D55E00 only for the loop icon.
Layout: four icons in a 2 by 2 grid with generous spacing, each centered in its own invisible square. No frames, no labels, no captions, no text, no watermark.
Background: perfectly flat solid #00ff00 chroma-key background for background removal. The background must be one uniform color with no shadows, gradients, texture, floor plane, or lighting variation. Do not use #00ff00 anywhere in the icons. Keep icons fully separated from background with crisp edges and generous padding.
I find this genuinely thrilling, and not for the hype.
Getting real work out of these systems is itself a craft, and that craft has matured in four waves as the underlying models grew stronger.
The waves are worth knowing because they explain how a bare chatbot became something that can drive a climate model, and they show exactly where a tool like the SUEWS-agent sits.
This is also, honestly, the story of why I got hooked and kept digging.
First came prompt engineering: with only the brain to work with, we learned to phrase the request carefully so it answered well, much like the plain-English neighbourhood description from the first post.
These waves overlap more than the order suggests, but it helps to take them one at a time.
Then context engineering: curating what the brain is given at each step, the right contents of the paper, so its answers are grounded in the real problem rather than generic.
In the SUEWS-agent that means feeding the brain the model’s schema, your site’s configuration and the readiness-check results at the moment it needs them.
Then harness engineering: giving the brain not just tools to act with but feedback signals, checks and validators, so it can catch its own mistakes and correct them rather than merely describing what it would do.
The SUEWS-agent is the SUEWS-specific part of that harness: the tools a host agent environment loads and uses to drive the model; the brain and the loop come from the host.
And most recently, what is now being called loop engineering: defining a goal worth pursuing and letting the agent work toward it, persistently, round after round, checking its own progress until the goal is met, and knowing when to stop if it is not.
This is the part that grips me as a researcher, because good research has always been less about the doing and more about asking a question worth answering.
Engineering a loop that keeps striving toward a good question is, to me, the most interesting thing happening in this field.
It is also exactly what we are testing at the hackathon: point the SUEWS-agent at a real urban-heat question and let it iterate toward an answer worth having.
I will come back to it later in the series.
The host environment I keep mentioning comes, in practice, in two main flavours, and they suit different people.
Claude Code (from Anthropic) is powerful and what I use daily, but it lives in the terminal, the black window with the blinking cursor, and was built for people already at home there.
Codex Desktop App (from OpenAI) is the gentler door I mean here: a familiar chat box that anyone who has used ChatGPT can drive.
Codex also has a command-line interface, closer in feel to Claude Code, but that is not the entry point I would recommend first to a non-programmer.
So my advice to a colleague with no programming background is simple: start with the Desktop App.
Either one can load and run the SUEWS-agent, the choice is not permanent, and the concepts above carry over completely.[1]
Everything in this post, the brain, the pen, the paper, the harness, the four waves of engineering, is general; it is true of any capable agent.
What makes it concrete for me is the SUEWS-agent: the SUEWS-specific harness layer we built for an urban climate model, with the readiness checks and validation that keep a general-purpose agent honest.
In the next post I’ll open it up and show how it is designed, and why those guardrails matter more than raw cleverness.
Claude Code began as a terminal-native research preview and became generally available in May 2025 after positive feedback, according to Anthropic’s Claude 4 announcement.
That terminal user interface still feels like something born from heavy internal engineering use: surprisingly powerful if you live in code, but not the friendliest starting point for a general user.
The business side is a nice anecdote too: Anthropic later said Claude Code’s run-rate revenue had grown to over US$2.5 billion, more than doubling since January 1, 2026, in its Series G funding announcement.
I use both Claude Code and Codex Desktop App; they have different temperaments, and I genuinely like both. ↩︎