Building Multiple Agentic AI Trading Portfolio Pods

By Kristian Fagerlie · 2026-06-15 · 6 min read

The shift that has changed my agentic AI trading this year isn't a smarter model — it's running many small, independent trading pods instead of obsessing over one perfect strategy. Today I want to walk through how I think about these pods, the exact workflow I use to spin a new one up with Claude Fable 5 and Codex, and two real examples: a Polymarket maker pod that's quietly in the green, and a stock pair-trade I built yesterday.

Watch the video:

What a trading pod actually is

A pod is a single, self-contained trading setup that runs in its own instance and minds its own business. One pod might be my Polymarket 5-minute up/down maker. Another might be a mean-reversion trade on a pair of correlated stocks. Another could be a SpaceX short. They aren't related to each other — each one runs on its own, with its own logic, its own data, and its own monitoring.

Some of these pods will lose money. That's expected. The point isn't that every pod wins — it's that when you combine ten or twenty uncorrelated setups, the portfolio as a whole ends up green. One loses a little, two gain a little, and the sum is what matters. It's the same diversification logic behind an index fund, except each "holding" is an autonomous agent I've built and tuned myself. For the rest of the year, my plan is simply to stack up as many of these pods as I can that look profitable on their own, then watch what the combined number does.

Why many pods beat one perfect strategy

The real reason I run a portfolio of pods isn't returns — it's discipline. If you only run one setup, you stare at it. You get impatient. You start tinkering, forcing trades, adjusting parameters to make it "do something," and you quietly destroy your own expected value. I've done it, and most people trading manually do it constantly.

With twenty pods running, no single one is worth obsessing over. You let each do its thing in the background, the same way you don't check an index fund every hour. The more setups you have, the less you interfere with any one of them — and less interference is usually the whole edge. This is the same emotional-discipline problem I hit when I built the agentic trading heartbeat: the hard part isn't the model, it's keeping your hands off it.

The workflow: idea → data → analysis → pod

Every pod starts the same way. I have an idea I want to check, and the first thing I do is get Codex to help me find the best source for the data I'll need. It can go either direction — sometimes I start from the data I think I need, sometimes I sketch the model first and then hunt for data to feed it. Either way, the data is the gold. Without good, fresh data, none of the AI matters, which is exactly the point I made in the data pipeline post.

From there the loop is short: pull the data, hand it to Claude Fable 5 to analyze, and if the numbers look good, build the pod and run it. Codex 5.5 on medium does the grunt work of locating sources and grabbing data; Fable does the heavier reasoning over the results. For the stock examples below, Codex pointed me at QuantConnect, which suggested the yfinance Python package — and a few minutes later I had five years of daily closing prices to work with.

Pod #1: the Polymarket 5-minute maker

The pod I built yesterday is another run at the Polymarket 5-minute up/down maker side, this time wired up with Claude Fable 5. It's a low-frequency pod — it doesn't fire often — but when it does act, it's been doing well. After about 24 hours it had made only 18 fills and was sitting around $76 in the green. Today was dead quiet, almost no movement for six or seven hours.

And that's the feature, not the bug. A slow, quiet pod is exactly what you want to be able to ignore. If this were the only thing I was running, I'd get restless and start trying to make it more active — and I'd probably wreck it. As one of several pods, it just sits there making its small, patient edge. If you're newer to this side of things, my Polymarket trading bot walkthrough covers how a maker setup like this comes together.

Pod #2: a mean-reversion pair trade

The setup I looked at this morning is a classic pairs trade — find two correlated stocks that historically drift apart and snap back together, then bet on the snap-back. My first instinct was Coca-Cola and PepsiCo, since they're both in the beverage world and feel related. So I pulled five years of closing prices and asked Fable to find reverse-to-the-mean opportunities on the two.

The answer was honest and useful: the clean mean-reversion trades happened in 2021–2023, when the Coke/Pepsi ratio was stable and correlation was high. But after some GLP-1 headwinds for PepsiCo, the two aren't really correlated anymore — not a great pair right now. So I asked Fable how to find pairs that can run this trade, took its criteria back to Codex, and had it surface candidates. One, the "VMA" pair, had very strong correlation over the last year. I grabbed that data, ran it through Fable, and it was a completely different animal: across 21 trades there were 15 winners and 6 losers, with the best signal — short the leader, long the laggard — worth around 4%. That's a pod worth building.

The rubber-band twins (how I actually learn this stuff)

Before I build a pod, I make the model explain the trade to me in plain language — because I don't want to ship code I don't understand. I asked Fable to explain the mean-reversion signal with an analogy, and it gave me one I really liked:

Imagine V and M are identical twins who go to the gym together every day — same workout, same food, same life. For five years they've always stayed within a few steps of each other. Now picture a rubber band connecting them. On normal days it's relaxed and they walk in lockstep — no trade. Then something happens and one twin speeds up. The band stretches. A little taut? Just watch. Really stretched? That's your signal — history says it always snaps back. Dangerously tight? Either it snaps back hard, or something has genuinely changed and the band is about to break — so you exit either way.

That's the whole trade in one picture: when the band is stretched, you short the twin who ran ahead and buy the one who lagged, betting they converge. Talking the models through ideas like this is half of why I enjoy building these pods — I learn something almost every day, and that learning sparks the next idea. Understanding the rubber band on stocks is exactly what made me wonder whether I can find a mean-reversion signal on Polymarket too.

Handing the monitoring to agents

Once a pod is live, I don't babysit it — an agent does. For the Polymarket maker pod I have a cron job that runs every two hours, checks that everything is healthy, and confirms it's still running fine. Fully autonomous. That's what makes the pod model scale: if monitoring twenty setups meant twenty browser tabs and twenty nervous humans, it would fall apart. Instead each pod gets a small agent watching its vitals, and I only get pulled in when something actually needs me.

None of this is revolutionary on its own. But now that we have models like Fable, Opus, and Codex that can find data, reason over it, explain it, and monitor the result, running a whole portfolio of independent pods finally feels practical for one person. That's the experiment for the rest of the year — stack the pods, let them run, and see if the combined number stays green.

Resources

• Hyperliquid — perps exchange I use for several trading pods
• Polymarket — prediction market for the 5-minute maker pod
• QuantConnect — where Codex pointed me for sourcing the stock data
• yfinance — Python package for pulling historical closing prices
• AI_automata Discord — where a lot of these pod questions come from