Constrain the blast radius of model mistakes, reduce token overhead, and execute real-world actions safely by moving state, execution, and authority out of the model.
| Model-Centric Agent Architecture | Invariant Architecture |
|---|---|
| Prompt carries execution state | Durable runtime state |
| Growing conversation history | Projected context |
| Large static tool surface | State-aware Runtime Actions |
| Model reconstructs progress | Committed progress is stored as facts |
| Tool calls become authority | Runtime validates authority |
| Opaque reasoning loops | Structured execution traces |
| Retries may repeat work | Durable intent & explicit idempotency boundaries |
"Invariant moves the parts that must be reliable out of the model and into infrastructure."
Stop making the LLM responsible for remembering state, reconstructing progress, deciding execution authority, carrying application context, and coordinating side effects.
The model retains what it excels at: semantic reasoning.
Invariant absorbs the implemented guarantees for state, execution, authority, persistence, and coordination primitives. Cross-process recovery orchestration remains outside the public Beta contract.
Probabilistic
│
▼
Semantic Reasoning
│
▼
Runtime Action
│
─── validated boundary ───
│
▼
Runtime
│
┌────────────┼────────────┐
▼ ▼ ▼
Workflow State Effects
execution commits dispatchSession → what carries forward (identity & hydrated application context)
Projection → what a consumer sees (pure derived views, PII redaction)
Agent → what should happen (model-driven proposals over workflows)
Workflow → what may happen (allowed paths & side-effect boundaries)
Capability → what touches the world (stored intent + idempotent effects)
Runtime → what makes it durable (event log, atomic commits & OCC)$$\text{ADMIT} \longrightarrow \text{EXPOSE} \longrightarrow \text{PROPOSE} \longrightarrow \text{COMMIT} \longrightarrow \text{EFFECT}$$
// 1. Projection: Only expose what the consumer needs (No PII)
const supportProjection = app.projection("support-view", ({ session, runtime }) => ({
customerTier: session.context.accountTier,
validActions: toTools(runtime.validActions),
}));
// 2. Workflow: Guarded transitions with transactional capabilities
const refundWorkflow = app.workflow("refund", {
inputSchema: RefundInputSchema,
})
.capability("load-order", loadOrderCapability)
.step("check-policy", evaluateRefundPolicy)
.branch("decision", ({ state }) => (state.eligible ? "APPROVED" : "REJECTED"), {
APPROVED: issueRefundFragment,
REJECTED: rejectRequestFragment,
});
// 3. Agent: Model proposes; Runtime validates and executes
const supportAgent = app.agent("support-agent", {
instructions: "Help customers resolve refund inquiries via registered workflows.",
projection: supportProjection,
workflows: [refundWorkflow],
});The runtime owns truth. Projections control exposure. Runtime Actions control authority. Capabilities deliver effects.