Every quarter, a new "AI trader" shows up on Crypto Twitter. The pitch is always the same: paste a chart and a few headlines into a prompt, get a recommendation back. Six weeks later, it's quietly forgotten.

It's not that AI can't help with markets. It's that the wrong shape of AI can't.

This is a long-form post about that shape — and why we built xris as a multi-agent system instead of a single LLM front-end.

Failure mode 1: hallucinated levels

Ask a frontier LLM where Bitcoin's nearest support is. It will give you a number. Confidently. With reasoning.

The problem is that the number was generated, not measured. The model has read about S/R levels. It's seen thousands of charts described in text. So it pattern-matches: "Bitcoin around $80k, support is probably $77k, that looks reasonable." It's a plausible-sounding number, statistically anchored to its training data.

But it's not anchored to the actual price action of the last 60 days. The model can't see that price touched $78,700 five times with volume confirmation, and it certainly didn't compute the cascade. It's vibing.

Our S/R agent, by contrast, runs a deterministic algorithm. It scans candles. It clusters touches. It checks volume confirmation. It runs an ATR-based cascade that filters levels too close to the entry. The output is a number you can trace back to the actual data. You can disagree with the algorithm — but you can't accuse it of hallucinating.

Failure mode 2: flattened context

A single LLM has one context window. Whatever you stuff into it competes for attention with everything else.

Try this prompt: "Here's the BTC chart, here's three pieces of news, here's the macro calendar, here's the fundamentals, what's the trade?"

What you get back is a response that weights all of those equally — because that's how attention works in a single forward pass. The unlock event two days from now sits next to the random Reddit headline and gets roughly the same weight in the synthesis.

A team doesn't work that way. The News agent has only read news today. It has its own ranking — pinned vs recent, high-impact vs low — built up over the day. When that News agent surfaces "$400M unlock in 36h," it does so with the full context of having seen, weighted, and dismissed the noise. It doesn't have to compete in the same context window with technicals.

Each agent's domain becomes a specialist's working memory, and the merge layer reads from those specialists rather than re-deriving everything in one pass.

Failure mode 3: no depth in any one domain

Frontier LLMs are generalists by design. They've read about technical analysis, on-chain analysis, macro, fundamentals — but they've read at the surface level for each. They know what RSI is. They probably can't compute a multi-timeframe regime correctly, can't keep up with the actual on-chain flows, and have no live access to FOMC dates.

A specialized agent doesn't pretend to know everything. It knows its domain deeply because it has the right tools, the right data sources, and the right primitives.

Our Fundamentals agent doesn't read about supply curves in a textbook — it pulls real-time circulating supply, FDV, max supply, vesting schedules, and computes the FDV/MC ratio with bands ("low overhang" / "moderate" / "high overhang"). When it tells you a token has 18% of its supply unlocking in the next 30 days, that's a primary number it just looked up, not a hallucinated estimate.

Depth comes from specialization plus tooling. A single LLM with no tools can't get there. A single LLM with all tools tries to do everything and does most of it badly.

Failure mode 4: no auditability

When a single LLM gives you a recommendation, you have no way to ask: which inputs drove this? What did you check vs invent? If I disagree with you, where exactly does our reasoning diverge?

This is the killer for serious capital. A trader who has had a position go against them needs to know why their model thought what it thought. Otherwise they can't update.

In a multi-agent system, the trail is built in. The S/R agent surfaces the levels with their tier and touches. The News agent surfaces the items with their sources. The Fundamentals agent surfaces the metrics with timestamps. When the merge layer combines them and prioritizes, you can click back to each agent's raw output. The system is legible by construction.

If xris flags a setup and you take the trade, and it goes wrong, you can ask: "What did the Risk agent say at entry?" Maybe it warned about a vol regime shift you didn't pay attention to. Or maybe nothing pointed at the failure, and you've learned that this category of move isn't currently observable. Either way, you've improved.

What multi-agent doesn't fix

To be precise: multi-agent doesn't make agents smarter at their individual job. The S/R agent is only as good as its algorithm. The News agent is only as good as its sources and ranker.

What multi-agent fixes is the synthesis problem. The "how do you combine inputs of different types, frequencies, and reliabilities into one decision-ready view." That problem is structurally different from the "how good is each input" problem, and a single LLM is built for neither.

A team is built for the second.

What we ship

xris runs five specialists today: technicals, news, fundamentals, macro, risk. Each is published. Each is auditable. Each disagreement gets surfaced rather than averaged into a confident lie.

If you've burned cycles on single-LLM trading copilots and gone away unimpressed, we think you'll feel the difference within a session. The dashboard is free during beta and you don't need an account.