Academy · Solutions · Agentic System

The Agentic System pattern

In one line. A governed domain brain: many collections and a fleet of agents powering a whole portfolio of use cases — with governance, compliance, recommendation and reporting operating over all of it at once. You’ll be able to. Recognize when you’ve outgrown single workflows, and stand a System up in the right order: domain model first, fleet second, workflows third, facets last.

What it is

An Agentic System is level 3 of the ladder. A Workflow automates one process; a System runs a domain — loan origination, compliance operations, a lending portfolio — where many processes share the same data, the same rules, and the same accountability.

Choose this shape when the workflows start overlapping: three processes reading the same collections, the same field meaning different things in different builds, auditors asking questions no single dashboard answers. Those are the signs you need one brain, not more arms.

The canonical architecture

A System is a layered stack. Read it bottom-up — that is also the build order:

 L5  OUTPUT & DELIVERY   decision package + provenance, findings     ┐
 L4  HITL                confidence gate; exceptions to a human      │ cross-cutting facets
 L3  VALIDATION & RULES  within/cross-doc checks, screening, flags   │ watch every layer:
 L2  AI INTELLIGENCE     classify, extract, transform, confidence    │ governance, compliance,
 L1  INTAKE              collection ingestion fires the pipeline     │ recommendation, reporting
 L0  PREPROCESSING       ids, hashing, PII pre-screen, page split    ┘
 ──────────────────────────────────────────────────────────────────
 THE SHARED DOMAIN MODEL — many collections + one taxonomy.
 Every agent reads it; every workflow writes to it; every facet reports on it.
 Worked by a FLEET of agents + shared skills.
  1. The shared domain model first. The collections and the taxonomy — the field dictionary every agent and rule speaks. One model, reused everywhere. Get this wrong and every layer above inherits the mistake. (D1, D3)
  2. A fleet, not a hero. Not one agent but many, each expert on a slice — a classifier, an income analyzer, a screener — sharing skills so common logic lives once. The reference lending build runs 50 agents and 18 shared skills.
  3. Workflows on top. Each layer’s process is a level-2 workflow — a mesh or XFlow moving work L0 through L5. Everything you learned in the Workflow pattern applies per layer.
  4. Facets over everything. Governance, compliance, recommendation, reporting.

The idea that makes Systems tractable: a facet is not a separate app — it is a lens over the same domain model. Compliance is the validation rules, the screening agents and the audit trail, taken together. Reporting is dashboards over the telemetry those same agents emit. You build one domain model and turn on four lenses; you do not build four systems.

The facets

Facet What it is Built from
Governance Who may touch what, versioned and auditable Access & roles, versioning
Compliance Rules, screening, and an immutable record Policy & guard rails, Audit
Recommendation Agents that score and decide, with provenance Decision agents (A4)
Reporting Dashboards and cost telemetry over everything Dashboards & inbox

Standing one up without drowning

Do not build all six layers on day one. The incremental path that works:

  1. Ship one workflow end to end on the shared domain model (weeks, not months).
  2. Add the second workflow on the same collections — resist the copy-paste fork; this is the step where the domain model earns its keep.
  3. Turn on the facets one at a time, compliance first if regulators are in the room.
  4. Package and version the whole thing when it stabilizes (B10 · Package & ship).

Worked builds

Build Teaches
B8 · Compliance evidence map The compliance facet standing alone — rules to evidence to audit
B9 · Capstone: Lending domain brain The full stack: domain model, fleet, layered workflows, all four facets

Where to go next

Prefer learning inside the product? The same academy lives in the platform's Learn menu — every screen links to the chapter that explains it.

See the platform live