How the AI analyst thinks
The agent reads the dealer book every 10 minutes and outputs a structured trade thesis. Here's what it's actually doing.
The agent on SPX-Flow gets called the “AI analyst” because it produces a structured trade thesis in plain English every 10 minutes. The naming is accurate — it learns from real options flow, classifies the regime, picks the levels, sets a flip line, and writes the call. Here’s a walk-through of what it does so you know how to read its output.
What it is
The agent ingests the same five metrics you see on the dashboard — NET POSITION, GEX, DEX, VEX, CEX — plus price action and time-of-day context. It outputs a structured object every 10 minutes: regime call, confidence, R/M/S levels, flip line, key drivers, and a one-line thesis. The structured fields are what you trade against; the prose is the human-readable wrapper.
How you use it
The agent banner sits at the top of the dashboard. Read it first: regime + confidence + levels + thesis + flip line. That’s your starting hypothesis for the session. Then check it against your own read of the metric stack below. They should agree most of the time. When they don’t, that’s information — usually the agent has caught a transition before you have, or you’ve spotted a confluence the agent has weighted lower. Use the agent’s confidence as a sizing input: 80%+ act, 50-80% wait or trim, below 50% pass. The R/M/S convention: R = resistance / call wall, M = magnet, S = support / put wall.
Differentiated advantage
The agent produces a structured call every 10 minutes — not a once-a-day take, not a trader newsletter, not a chat-based oracle. It runs through every session continuously and learns from the SPX flow it’s been trained on. As more data accumulates, the calls sharpen.
Why you should care
A second pair of eyes that never gets tired, never holds a grudge from yesterday’s losing trade, and prints a structured thesis every 10 minutes is genuinely useful. Read the agent as a co-analyst, not an oracle: fast, consistent, complementary to your own read. The next two articles dig into the regime classifier and the confidence score so you know what each part of the call actually means.
The regime classifier, explained
The classifier reads the dealer book, the price tape, and the clock — then calls one of five regimes with a confidence.
The regime classifier is the core engine inside the SPX-Flow agent. It runs every 10 minutes during market hours, reads the current state of the dealer book and the price tape, and emits a structured regime call that everything else builds on. This article walks through what it sees and what it outputs.
What it is
The classifier takes four inputs: gamma regime (net long, net short, or balanced — read off GEX aggregate and NET POSITION cluster), market regime (whether second-order forces from VEX and CEX are dominant or quiet), displacement state (how price is moving — coiling, drifting, trending, chopping), and price-action confirmation (whether recent moves match what the gamma + market regime would predict). From those four it picks one of five regimes: Pin, Squeeze, Trend, Grind, or Transitional.
How you use it
Each regime maps onto a specific combination of the four inputs. Pin = net long gamma + quiet second-order + low displacement + range-bound price. Squeeze setup = net short gamma near spot + building DEX + coiling price + failed-break pattern. Trend continuation = capitulated book with no nearby walls + supportive second-order + persistent direction. Grind = net long but diffuse gamma + quiet + low displacement + random price. Transitional = inputs disagree, no clean fit. When the classifier returns Transitional, that’s the system being honest — two regimes scored similarly and it won’t pick one. The right play in Transitional is “wait.”
Differentiated advantage
The classifier evolves with the data. As SPX-Flow accumulates more sessions, the regime templates and weighting refine. Replays get re-scored against the current logic so you can see how the latest version handles past sessions. A classifier locked to a static rulebook would stale; one that learns from real flow gets sharper over time.
Why you should care
When the classifier prints a regime with high confidence, four specific input conditions agreed. When confidence drops, the inputs disagree. Reading the regime as the output of a known process, not a pronouncement, lets you calibrate trust over time. The next article unpacks how the agreement gets quantified into the confidence number.
Confidence scores
The 0-100% confidence on every call is computed from how strongly the metrics agree. Here's how to use it for sizing.
The 0-100% confidence on every agent call isn’t a vibe — it’s a computed value that captures how strongly the metrics agree. Treat it as a sizing input. The dashboard publishes the number on every call so you can build operational discipline around it: this many points = act, this many = wait, this many = pass.
What it is
Confidence is built from how cleanly the classifier’s inputs agree with the chosen regime template, how tight the geometric structure of the call is on the strike chain, and how fragile the levels near spot look. When NET POSITION, GEX, DEX, VEX, and CEX all point at the same regime with crisp R/M/S levels and no decaying walls nearby, confidence is high. When inputs disagree, structure is diffuse, or a key wall is weakening, confidence drops.
How you use it
The thresholds map onto operational behavior. 80% and above: inputs agree, structure is clean, no fragility concern. Act on the call. Size normally. Manage against the published flip line. 65-80%: mostly aligned with some tension. Reduce size, wait for one additional confirmation, or trade tighter. 50-65%: inputs are mixed. Treat as informational — narrow your read of the day, don’t enter on it. Below 50%: ignore. Either the regime is genuinely transitional or no clean read is possible. Don’t fight the data. Confidence updates every 10 minutes alongside the regime call. If you’re in a trade based on an 84% Pin and the next snapshot prints 71%, the regime is still Pin but something tilted weaker. Read the updated thesis to see what.
Differentiated advantage
Many flow products publish only the call, or compress confidence into a coarse “high/medium/low.” SPX-Flow publishes the precise number because it changes how you size. A 79% Pin and an 84% Pin produce the same regime label but warrant different position size. Hiding the number makes the product feel more authoritative; surfacing it makes it more useful.
Why you should care
Confidence is a sizing input, not a permission slip. The discipline is binding your position size to the number — bigger when the metrics agree, smaller when they’re tense, none when they’re mixed. That single habit prevents most of the “I knew it was 60% but I went big anyway” losses retail tends to take.
How the AI gets sharper
The agent learns from real SPX flow. As the data accumulates, the regime calls tighten and the edge compounds.
The agent isn’t a one-shot product. It’s a system that gets sharper as it sees more SPX flow. Every session it processes adds to the data the calls are built on — and the calls themselves are open enough to inspect, so you can develop your own intuition for when to follow them and when to override.
What it is
The agent reads the dealer book and price tape every 10 minutes, classifies the regime, picks levels, and prints a thesis. Underneath the call is a structured object — the regime vote, the confidence components, the source strikes for each level, the gatekeeper for the flip line. That structure is what you inspect when you want to know why the call came out the way it did. Each call is reproducible against the metric snapshot it was generated from, so any past call can be re-examined alongside the data it saw.
How you use it
Three workflows. Open the structured object below the prose thesis — regime vote, confidence components, levels with source strikes, flip line with gatekeeper. Stack the metric snapshots from that timestamp in replay so you see exactly what the agent saw. Trace the votes to find the swing metric — the one whose vote tipped the call. After ten or twenty inspections, you build a mental model of when the agent is most reliable for your trading style. Some traders find Pin calls highly reliable but Trend calls weaker for their setup, or vice versa. The audit trail is what lets you discover that.
Differentiated advantage
The agent is built to keep getting sharper. As more flow data accumulates and regime patterns refine, calls tighten and confidence calibration improves. Replays get re-scored against the latest version so you can see how today’s logic would have read past sessions. That trajectory — system, data, calibration, repeat — is the long arc of the product.
Why you should care
Tools that hide their reasoning create users who never improve. Tools that let you inspect create users who get sharper alongside the system. Three months in, a user who’s been reading the structured calls regularly will have a more sophisticated read of dealer flow than they had at the start. That compounding is the trust model the dashboard is built around — not “the AI is right,” but “here’s what it saw, here’s what it called, judge for yourself.”