Chasing Market Data with Hyperagent
This is the research assistant that doesn't mind the grit.
Ranch Systems started as a newsletter about ranching operations and the tools that make them work. It still is. But it's also about the people, the culture, the failures, and the weird tangents — the stuff that actually matters when you're building something real in rural America. Some weeks that's Airtable. Some weeks that's why venture capital doesn't understand the West. Some weeks it's a story about getting thrown out Wizzards bar in Borger with a group of Hells Angels, and what that taught me about personality.
What Happened When I Turned an AI Agent Loose on the CME
I’ve been building a crush margin engine for a cattle feeding client — a structured system in Airtable that tracks the three numbers that determine whether a feeding operation makes money: feeder cattle in, fat cattle out, and corn in between. The analytical framework was solid. What I didn’t have was the raw fuel: a full year of daily market data from the CME for all three commodities.
Pulling 750 data points by hand — settlement prices, daily ranges, and volume for Live Cattle, Feeder Cattle, and Corn across every 2025 trading day — wasn’t something I was going to do manually. And I knew from experience that the CME doesn’t exactly hand this stuff out. So I decided to throw a new tool at it: Hyperagent.
Talking to the Agent
Hyperagent is a browser-based AI agent — you tell it what you’re after in plain English, point it toward some resources, and let it work. That’s essentially what I did. I gave it a prompt that Claude had helped me refine based on the crush margin project, pointed it toward a few links, and told it what I needed: daily futures data for LE, GF, and ZC across the full 2025 trading calendar, structured for Airtable.
What happened next surprised me.
The agent didn’t just start scraping. It asked questions. Good ones — some I could answer with a checkbox, others that required me to think. Those answers led to deeper questions, and those to deeper ones still, until the agent had built a real understanding of what I was after and why. It felt less like configuring a tool and more like briefing a sharp research assistant who happens to work at machine speed.
Then it went to work.
The Hunt
I watched it navigate the CME DataMine API — and hit the same wall I would have. Paid credentials required. Door locked. It tried the CME’s public settlement API and found what I already suspected: only about a week’s worth of data at any given time. A window, not an archive.
So it pivoted. It dug into Yahoo Finance, Barchart, and a handful of other data sources I hadn’t even thought to check. It evaluated each one, compared coverage and reliability, and settled on Yahoo Finance’s futures feed via a Python library called yfinance — which delivered clean daily data for all three commodities across the entire year. Settlement price, open, high, low, volume. Exactly what the engine needed.
The one thing it couldn’t find anywhere for free: open interest. That data lives behind paywalls across every source it checked. No free provider on the internet offers daily historical open interest for CME futures. The agent noted the gap, left the column in the table structure, and moved on. Smart.
Building the Table
Here’s where it got really interesting. Hyperagent located the Airtable base I’d been working in — the one that already held market snapshots, performance benchmarks, and hedge scorecards — and built a new table inside it. “CME Daily Data 2025,” with fields for date, commodity, ticker symbol, settlement price, the full daily range, volume, and that placeholder for open interest. It asked a few follow-up questions about the experience, whether it should store the skills it had developed for future use, and what steps were next.
The whole process — from initial prompt to 750 structured records ready for import — took a fraction of the time it would have taken me to do manually. Months of work compressed into a single session. And I didn’t write a line of code.
What Worked, What Didn’t
What worked: the conversational approach. Hyperagent’s question-asking phase was genuinely useful — it forced me to articulate requirements I might have left vague if I’d been doing this myself. The agent’s ability to evaluate multiple data sources, hit dead ends, and pivot without me having to redirect it was impressive. It handled the CME paywall the same way I would have — tried the front door, tried the back door, then found a better route.
What didn’t: open interest remains the gap. No agent can pull data that doesn’t exist in a free, accessible format. That’s not a tool limitation — it’s a market data reality. When CME DataMine credentials come through, the architecture is ready.
From Data to Vision
Once I had the data — all 750 records in a clean CSV — I brought it to Claude and asked for charts. Not just any charts. I wanted to see a year of cattle and corn markets the way I think about them: layered, comparative, revealing the patterns that matter for crush margin analysis.
What Claude produced was something else entirely.
I’ve worked with data analysts. I’ve seen what comes out of Tableau and Power BI. The charts Claude built — in seconds, not weeks — were some of the most beautiful, most readable market visualizations I’ve ever seen. Settlement price trends for all three commodities, volume overlays, daily range analysis. The kind of work that would normally require a skilled analyst and a top-notch designer working together over months. Claude just does it. No fuss, no twelve-round revision cycle, no “can you move the legend two pixels to the left” back-and-forth.
For someone like me — someone who processes information visually — this was the moment the whole workflow clicked. The data was real. The structure was sound. And the visualization made it all make sense at a glance.
What This Means
This project started as a data collection exercise. It turned into something bigger — a proof of concept for how AI agents and AI assistants can work together in a real consulting workflow. Hyperagent did the hunting and building. Claude did the refining and visualizing. I did the thinking and directing. Each piece played to its strength.
The crush margin engine has its fuel now. The analysis has begun. And the toolchain that got us here is something I’ll be using on every project going forward.
Claude Charts
The charts below were developed entirely in within Claude.ai using a .skill that provides guidance about my brand guidelines. If you’re on mobile, the charts are scrollable. This is some really, really cool stuff! 🔥
If your information is scattered across spreadsheets, apps, and hard drives, it’s time to lean out your tech. I provide a structured Systems Audit to eliminate software bloat and build a digital foundation that actually works for your operation.
About the Author
This post was written by Walker Milhoan, founder of Milhoan Design. He spent fifteen years working ranches from Texas to Montana — day work, horseshoeing, sale barns — before building ranch management software twice and failing both times. Now he builds systems for working operations and writes about what he learns: ranching, web development, AI, data, design, and the culture of rural business. Sometimes those stories overlap. Sometimes they don't.