The Shared Brain.

I am building a py-native meta-EA: one engine that holds the whole book under a single roof, with a regime filter and risk manager deciding which strategies are allowed to trade.

Right now I run my book as a set of separate automated strategies. They all sit on MT5, each minding its own logic, each sized on its own rules. It works. It's also a faff to manage, and it leaves something on the table.

So the thing I'm building at the moment is a py-native meta-EA: one engine that holds the whole book under a single roof, with a regime filter & risk manager sitting on top deciding which strategies are allowed to trade in the conditions we're actually in.

This is a build log, I'm partway in and some of my features are still left to build. Here's where it's at.

The why

Running strategies as separate EAs means the portfolio has no shared brain. Each system trades whenever its own signal fires, blind to what the others are doing and to the regime the market's in.

A trend strategy will happily keep firing into a chopfest because it can't see the regime turned. A single engine can. Detect the regime once, centrally, and I can switch whole families of strategies on and off instead of letting each one learn that lesson the slow way through drawdown.

The hard part

Regime detection is the bit I'm still chewing on. It's easy to build a filter that looks great in hindsight, because in hindsight every regime change is obvious.

The trick is a filter that calls the regime close to real time, without lagging so far behind that it switches strategies off right as the old regime ends. I'm testing volatility thresholds, correlation-structure shifts across the FX, index and metals blocks, and breakpoint-style detection. None of them is clean yet. Each one trades responsiveness against false signals, and I haven't found a balance I'm happy with.

Where AI actually helps

I've been using Claude & Codex simultaneously as build partners through this, and it's brilliant at some parts and useless at others. It's genuinely sharp at red-teaming logic: hand it the regime-filter rules and ask where they'd misfire, and it surfaces edge cases I'd otherwise have found three weeks later in live.

It's also quick for drafting scaffolding code. Where it falls down is judgement that needs the live book in front of it. It can propose twenty regime definitions. It can't tell me which one survives contact with my actual slippage and correlation structure. That call stays mine.

What I haven't figured out

The scope keeps moving. Do I let the meta-EA resize strategies dynamically based on regime confidence, or just gate them on and off?

Dynamic sizing is more powerful and far more dangerous, because a regime-confidence number that's wrong now controls position size across the whole book instead of one corner of it. I'm leaning toward gating first, sizing later, once the regime call has earned my trust. But that's my current gut-feel, not a final decision.

That's the state of it. One engine, one brain, a pile of open questions. I'll post the next installment when I have more updates on this.

Work in progress, back to it.