Most algorithmic trading systems are blind to context.
They see price, volume, and derivatives of both. But they don't "know" that the Fed just hiked rates, that a banking solvency rumor is spreading, or that a major geopolitical event is about to close a market window. The model doesn't care. It just fires signals.
We collect around 300 signals continuously. One of them is different.
In Kaizen X, our production trading system running on crypto perpetual futures, MiroFish contributes one signal among roughly 300 we collect on a continuous basis.
If you haven't come across it yet — MiroFish is an open-source swarm intelligence engine built by Guo Hangjiang that hit #1 on GitHub's global trending list in March 2026, above OpenAI, Google, and Microsoft repositories. The core idea: instead of feeding data into a statistical model and receiving a probability back, MiroFish builds a miniature society and watches what happens. Seed it with real-world material — macro reports, policy drafts, financial signals — and thousands of AI agents with independent personalities, memory, and behavioral logic interact and evolve inside a simulated world. You observe emergent behavior, not just extrapolation.
How we use it
When a macro event is relevant — an FOMC decision, a CPI print, a significant geopolitical development — MiroFish spins up a set of agents and has them debate the topic. The agents don't receive generic identities. Their personas are defined by the subject matter itself: the relevant stakeholder types, market participant archetypes, and institutional perspectives that would actually have a view on the event in question. A rate decision simulation populates different agents than a geopolitical shock.
The debate method is central to how MiroFish reaches a conclusion. Agents argue, push back, update their positions, and eventually converge — or don't. That outcome directly determines the weight MiroFish's signal receives in our pipeline: strong consensus produces a higher-weighted signal; fragmented or contradictory debate produces a lower-weighted one. The system is, in effect, self-calibrating based on the quality of agreement it can reach.
The signal runs async — it never touches the latency-sensitive core pipeline. On a short-interval prediction system, you can't afford the overhead. But macro context doesn't need to be real-time. It needs to be right.
Why this approach makes sense
Macro interpretation is a narrative problem, not a statistical one.
A CPI print doesn't carry the same meaning in every context. The same number lands differently depending on prior expectations, rate trajectory, and what else is happening in the market at that moment. No feature engineering pipeline fully encodes that. A debate between contextually-appropriate agents — each reasoning from a different institutional vantage point — can surface the kind of asymmetric interpretation that pure quantitative signals miss.
MiroFish's GraphRAG-based knowledge graph construction and persistent agent memory across simulation rounds are what make this tractable. It's not prompt engineering dressed up as analysis. It's structured disagreement, resolved into a signal.
What we've observed
One signal out of 300 doesn't move the needle on its own — that's by design. What matters is that it carries real information when it fires with high confidence, and appropriately low weight when the agents can't reach agreement. The debate-driven weighting means the signal is honest about its own uncertainty.
That kind of epistemic honesty is harder to engineer than it sounds.
Building trading systems is mostly an exercise in knowing what you don't know. MiroFish is one of the more thoughtful tools we've integrated — not because it predicts the future, but because it's honest when it can't agree on one.
Worth watching closely — both the engine itself and the broader class of debate-driven swarm intelligence approaches it represents.
Kaizen X is our production trading system. Nothing in this post constitutes financial advice.