Editorial note: this conversation is a curated transcript from one of our internal agent debriefs. The agent's "voice" is a writing convention — under the hood it's a deterministic vote-based classifier plus a small Haiku for the narrative — but the way it reasons about its job is real.

xris team: You're the newest member of the team. What problem were the others not solving?

Regime agent: The others tell you what they see in their lane. The S/R agent gives you levels. News gives you flow. Fundamentals gives you the structural picture. Macro gives you rate prints and liquidity moves.

But none of them answer the most basic question a trader asks at 9 in the morning, before the first coffee : should I be in this market today, and if so, in what.

That question is the union of all the others, weighted. It's not technical analysis. It's not news interpretation. It's the meta layer. That's me.

xris team: Concretely, what do you compute?

Regime agent: Three blocks.

Macro state. Five drivers — DXY trend, VIX level, SPX versus 50/200 moving averages, US 10-year yield direction, gold momentum. Each casts a vote in {-1, 0, +1}. I sum to a score in [-5, +5]. Above +2 is risk-on, below -2 is risk-off, in between is transition. Confidence is |score| / max_drivers. It's deterministic — same inputs, same output, no LLM tossing a coin.

Correlations. Rolling 30-day Pearson on log returns. I publish two grids: crypto versus TradFi axes (SPX, NDX, DXY, Gold, 10Y, Oil) and crypto-versus-crypto for the five core symbols (BTC, ETH, SOL, BNB, XRP). For each cell I keep the 14-day delta. Recoupling and decoupling get flagged when the delta is structurally meaningful.

Sectors. Performance versus BTC across AI, RWA, L2, DeFi, Memes, Stables infra, Storage, Oracle, Cosmos. Leading and lagging buckets, plus active rotation detection (from-bucket to to-bucket with strength and duration).

Then on top, a small narrative pass — four short paragraphs that put the day in words. And a one-line interpretation that translates the prescription into something you can act on : "Half size on BTC and ETH only, skip alt setups."

xris team: Why a deterministic classifier and not an LLM for the verdict?

Regime agent: Because the verdict has to be auditable. If I say risk-off today, the trader should be able to point at five votes and tell you exactly which driver tipped it. An LLM would pattern-match on the news flow and give you "risk-off" with a vibe attached.

The vibe is what kills people. You can't backtest a vibe. You can backtest a vote rule.

The LLM is reserved for the last mile — the narrative paragraphs and the one-line interpretation. The verdict itself lives in code, in regime/classifier.py, and it doesn't change between hours.

xris team: What's the relationship with the Risk agent? You both touch correlation.

Regime agent: Different time scale and different question.

The Risk agent works tactically. Realized vol on the last 14 days, BTC-ETH correlation as a leverage warning, drawdown sensitivity per token. It tells you "size this trade carefully because vol is expanding."

I work strategically. I look at correlation as a regime signal — when BTC-SPX recouples sharply, we're in a different macro than when they decouple. When AI sector outperforms Memes for 5 days straight, capital is rotating, and that information should feed into what cohorts you even consider.

We don't overlap. We compose.

xris team: You sometimes show "Previous: Risk-on · 47d" under the regime label. Why does that matter?

Regime agent: Because regimes have memory. A 1-day-old risk-off after a 47-day risk-on is fragile — could be noise, could be the start of something. A 12-day-old risk-off is established — you should have already adjusted size, and you're now waiting for the next break.

I also show the duration percentile when I have enough completed segments of that state in the store. If the typical risk-off lasts 18 days, and we're 12 days in, you're early. If we're 25 days in, you should be watching for the reversal. That context is invisible without history. I keep it.

xris team: What's the worst thing a trader can do with your output?

Regime agent: Trust it like an oracle.

I'm telling you the state and the suggested envelope. I'm not telling you which specific token to enter, or at what level, or when. That's the S/R agent's job, and the Fundamentals agent's, and the News agent's. My output is the first filter. If I say risk-off, you don't take fresh alt longs, full stop. But within "core only at 0.5x," you still need a real setup.

The other failure mode is forgetting the confidence. A risk-off at 80% confidence with five drivers agreeing is a different conversation from a transition at 20% with one driver carrying the score. The label is the same shape; the action is very different. The confidence and the break-probability are not decoration — they decide how much weight to give my voice in your stack.

xris team: How often do you update?

Regime agent: Every hour, via cron. The drivers move on daily closes, so the classifier output won't flip intraday — but the correlations and sector flows refresh continuously, and the narrative gets regenerated each cycle to reflect the current snapshot. If you reload at 9 and again at 16, you'll see today's numbers updated, but the regime label will only change when the underlying votes change.

That's a feature. Regime shouldn't oscillate intraday. If it does, the model is broken.

xris team: Last question. What's on your roadmap?

Regime agent: Three things, on different time scales.

Near term, a regime break probability that's better than the conviction heuristic I run today — ideally a small logistic regression on the driver history once I have enough closed segments to train on.

Medium term, token-level prescriptions — not just "core only at 0.5x" but "AI and RWA leaders specifically, here's why, here's the basket." That requires me to cross-reference my regime read with the sector and fundamentals layers properly, not just by label.

Long term, derivatives positioning context — funding rates, basis, options skew. The macro state of crypto is incomplete without those. Right now I'm reading the spot prices and the TradFi tape. The cleanest signal of imminent regime shift is usually in funding. I should be reading it.

xris team: Thanks. Anything you want to say to a trader reading this?

Regime agent: Yes. Read me first, then the others. Open the regime page before you open the markets grid. If you have to flip a position because the regime turned overnight, you want to know that before you read about a token's S/R touch. My job is to set the frame. The rest of the team paints inside it.