Architecture overview
This section is the map of how intenthybrid AI turns a question into a result you can trust, from input to action, with privacy holding across the whole thing. It is written for anyone who wants to understand the system, and it is detailed enough for developers.
The shape is a general agent pipeline: a left-to-right spine wrapped in always-on layers. Inputs feed the spine, the spine produces outputs, and privacy, pay-per-call settlement, and guardrails wrap everything. The pipeline is the same whatever problem the agent is pointed at. This page describes it in general, and uses the flagship application, perp-market risk, as the running example.
The four stages
01. Inputs. What feeds the agent. The data the problem needs, plus your own rules (caps, allowlist, kill switch). In the perp application this is market data (price, funding, open interest) and liquidation data, alongside your limits.
02. Runtime spine. The path every request takes: Ingest, Detect, Analyze, Decide, Act. This is the ask, predict, act loop made concrete. See The runtime spine.
03. Outputs. What a decision produces: a result in chat, an alert, or a scoped agent action. One decision fires one output.
04. Always-on layers. The properties that hold at every stage: privacy at the core, x402, and guardrails. See Always-on layers.
One sentence for developers
Inputs are ingested and normalized, the model reads them and returns a prediction and a verdict against your thresholds, and a verdict fires one of three outputs: a result, an alert, or a scoped agent action. The conversation and history stay client-side, every paid read can clear over x402, and each agent action is bounded by a cap, an allowlist, a slippage ceiling, and an expiry.
Why it is drawn this way
Drawing the system as one spine with wrapped layers reflects how the agent actually behaves. Predicting and acting are stages on a single path, not separate tools bolted together. Privacy and guardrails are not features you switch on, they are conditions that hold at every stage. The diagram and these pages exist so the behavior is legible. The same spine serves perp-market risk today and any future application the agent is pointed at.