Azirella

Research-backed architecture

Technology

Built on peer-reviewed research, enterprise data standards, and a decade of supply chain optimization practice.

The engine behind the Decision Stream: one shared world model, four tiers of coordination, and the AIIO operating model.

Four Tiers of Decision Coordination

Context, guardrails, and targets flow down as policy and directives. Feedback and outcomes flow back up. Each tier operates at its natural time horizon.

Strategic / Network

Strategic · Weekly / Monthly

Network-wide policy optimization. Sets guardrails, KPI targets, and policy parameters that shape all downstream behavior. Design, IBP, and S&OP decisions.

4

Tactical / Network

Tactical · Daily

Translates strategic directives into site-specific context. Generates demand forecasts, priority allocations, and cross-site trade-offs. Cross-functional negotiations resolve at machine speed.

3

Operational / Site Coordination (per site)

Operational · Hourly

Each site has its own coordination layer. Learns causal relationships between the site's execution agents, predicts cascade effects (a production spike will create quality load hours later), and pre-adjusts urgency before cascades unfold.

2

Execution / Site & Role (per site)

Execution · <10ms per decision

Each site runs its own hive of 11 specialized agents making narrow decisions at machine speed: ATP, purchase orders, rebalancing, manufacturing, quality, maintenance, and more. Each agent decides within guardrails set by the tiers above.

1

Remember: Velocity creates value.

Platform Architecture

Four tiers of agent-driven decisions. Context Engine routes external signals into each tier by horizon. Experiential Knowledge shapes agent priors forward and captures overrides for retraining back. The Digital Twin trains agents — it is not a planning surface.

Context Engine
Signal routing by horizon
Internal
External
Comms
Azirella
IoT
L4 Strategic
weekly · whole network
Policy Optimisation
Network-wide policy θ, KPI targets, guardrail envelope
L3 Tactical
daily · cross-entity
Domain-Model Reconciliation
Demand / Supply / Inventory / Capacity — 4-way lateral convergence
L2 Operational
hourly · single node
Node Coordination
Intra-site urgency modulation, MO disaggregation, always-on
L1 Execution
<10 ms · single decision
Atomic Decisions
11+ Execution Agent Hive, conformal-bounded
Experiential Knowledge
Priors forward · Overrides back
Capture
Classify
Route training
DATA MODEL · INTEGRATION · MCP
Context Engine — compiles strategic intent (strategy docs · OKRs · executive directives) into the cascading guardrails, KPI targets, and AIIO thresholds that constrain every tier. The intent-engineering layer most platforms force you to write yourself.
Digital Twin — training environment (virtual clock, scenario engine). Feeds L1 pre-training and L2–L4 provisioning. Not a planning surface.
Governance Pipeline — envelope → impact scoring → AIIO mode → guardrail directive. Cross-cutting across every tier.

Decision Intelligence Architecture

Autonomy is built as a Decision Intelligence Platform following Gartner's DIP framework (inaugural MQ, January 2026). The architecture implements all four DI lifecycle capabilities natively for supply chain.

DECISION INTELLIGENCE 01 MODEL Decision Modeling Five decision elements define every decision structure 02 ORCHESTRATE Decision Orchestration Agent coordination + AAP coordinate execution 03 MONITOR Decision Monitoring Drift detection, calibrated confidence, quality scoring 04 GOVERN Decision Governance Override tracking, confidence bounds, escalation arbiter

Architecture informed by decision science research - turning information into better actions, mapping decision levers through intermediaries to outcomes. Concept adapted from FlexRule's DI Architecture and Wikipedia: Decision Intelligence.

Platform Architecture (Detailed)

Four decision tiers cascade context, targets, and guardrails downward while feeding escalation and feedback upward. From strategic network design through tactical planning, operational coordination, and sub-second execution, every layer is connected by a unified data model, experiential knowledge layer, and integration fabric.

Strategic Intent & Operational Signals In
Internal Sources
Strategy · OKRs · Policies
External Sources
Weather · Sentiment · CPI
Communication
Email · Slack · Teams
Azirella Assistant
Executive & User Directives
IoT & Events
Sensors · Alerts · Triggers
Context Engine
Parse· Classify· Route by Horizon· Inject
Continuous · Multi-channel · Temporal
weekly daily hourly immediate
Strategic / Network · Weekly
Design· IBP· S&OP
Policies · Guardrails · Cross-functional alignment · Risk scoring · Portfolio optimization · KPI Targets
Tactical / Network · Daily
Forecast
ML Baseline
Demand
Shaping · Sensing
Supply
MPS · MRP · Sourcing
Inventory
Rebalancing · Buffers
Capacity
Shifts · Buffers
Operational / Site Coordination · Hourly
Cross-Function Trade-Offs · Urgency Modulation · Causal Coordination · Likelihood Calibration · Bottleneck Detection · Synergy Signals
Plant A
DC West
DC East
Plant B
DC Central
Execution / Site & Role · <10ms
Agent Hive per Site
<10 ms · A2A-native peer coordination
ORDER FLOW AATP PO OrdTrk EXECUTION MO TO Quality SubCon Maint. PLANNING ADJUSTMENT Rebal. FcstAdj Buffer
Experiential Knowledge
Genuine vs Compensating Classification
Override capture · Pattern detection · Agent training integration
State Augmentation Reward Shaping Likelihood Calibration Simulation Modifiers
Data Model
Multi-tenant · Sites · Products · BOM · Inventory · Supply Plans · Demand Plans · Purchase Orders · Sales Orders · Shipments · Vendors · Customers
Integration
Master Data · Transactional Data · Delta/Net Change Loading · Fuzzy Table & Field Matching · AI Schema Validation · Z-Field Interpretation · Auto-Fixing
Acme Co SAP
S/4HANA
Beta Inc SAP
ECC
Gamma Ltd D365
Dynamics
Delta Co Odoo
v18
Epsilon ORCL
Oracle
Zeta Corp LOG
Logility
Eta Inc KNX
Kinaxis
+
More
MCP · Agent Protocol
Model Context Protocol · Agents read masters and write decisions back as tool calls · Provenance preserved
PRIORITY D365 SCM
MCP server
any MCP server
extensible
Context, guardrails, & targets down Escalation & feedback up Context engine (5 input channels) Horizontal integration 4 decision tiers · 5 network agents · 11+ site agents Roadmap (not yet live)

Sequential Decision Framework

Autonomy's four-tier agent stack instantiates Warren Powell's Sequential Decision Analytics (SDAM) framework. Each tier uses the policy class that fits its horizon and scope. The Digital Twin is the training environment that pre-trains these agents — it does not generate plans.

STRATEGIC weekly · whole network S&OP · policy θ · KPI targets · guardrail envelope CFA — Cost Function Approximation TACTICAL daily · cross-entity Demand / Supply / Inventory / Capacity · 4-way lateral convergence CFA / VFA blend OPERATIONAL hourly · one node Node Coordination · urgency modulation · MO disaggregation · always-on VFA — Value Function Approximation (tenant-adapted) EXECUTION <10 ms · one decision 11+ Execution Agent Hive · conformal-bounded decisions VFA — Value Function Approximation (generic, frozen) SHORTER HORIZON · HIGHER FREQUENCY
Digital Twin — training environment. Feeds L1 pre-training and L2–L4 provisioning. Not a planning surface.
Governance Pipeline — envelope → impact scoring → AIIO mode → guardrail directive. Wraps every tier's decision.

Execution Agent Architecture

Each agent uses a compact, purpose-built architecture that achieves near-optimal performance while running at <10ms per decision.

INPUT Decision Context AGENT CORE FEATURE Extraction DECISION Model CONFORMAL PREDICTION Confidence Calibration Layer CAUSAL ATTRIBUTION Counterfactual Evaluator OUTPUT Decision + Confidence FEEDBACK Outcome Signal <10ms per decision near-optimal accuracy

Small, focused models that learn generalizable rules rather than memorizing patterns. This keeps agents fast, explainable, and reliable in production.

Conformal Prediction

Confidence Guarantees

Every AI decision carries a distribution-free confidence guarantee via conformal prediction. Unlike heuristic confidence scores that require distributional assumptions, conformal prediction provides mathematically calibrated coverage: the true outcome falls within the prediction set with guaranteed probability, regardless of the underlying data distribution.

This guarantee holds even when the model is wrong - misspecification widens the prediction set but never breaks the coverage guarantee.

95%+
Coverage guarantee
0
Distributional assumptions
PREDICTION SET (95% COVERAGE) Value Time Uncertainty grows = band widens = routes to higher tier Prediction Confidence band Actual outcome

Concept adapted from Awesome Conformal Prediction and BBVA AI Factory.

Distribution-Free

Coverage guarantee holds regardless of the underlying data distribution - no Gaussian assumptions needed.

Continuous Calibration

Runs continuously from historical decision-outcome pairs. Adapts to seasonal demand, supplier disruptions, market shifts.

Automatic Routing

When prediction set size grows, decisions automatically route to higher reasoning tiers for more thorough analysis.

COUNTERFACTUAL REASONING DECISION Point ACTUAL Outcome A Agent decided X COUNTER- Outcome B What if decided Y? CAUSAL IMPACT TRAINING Weight by impact OVERRIDES Measure effectiveness GUARDRAILS Expand / contract Actual path Counterfactual Causal impact

Concept based on Interpretable ML - Counterfactual Explanations (CC BY-NC-SA 4.0).

Causal AI

Decision Attribution

Knowing an outcome is good is not enough - you need to know whether the decision caused the good outcome. Autonomy uses counterfactual reasoning to evaluate every agent decision: what actually happened vs. what would have happened if the agent had decided differently.

This causal attribution drives everything downstream. Agent training weights decisions by causal impact, not outcome correlation. Override effectiveness is measured by comparing the human's outcome against the agent's counterfactual.

Without causal reasoning, an autonomous system reinforces lucky decisions and punishes skillful decisions that encountered bad luck. Causal AI ensures the system gets genuinely smarter with every decision cycle.

Agentic Authorization Protocol

When an agent needs to take an action outside its authority boundary, the AAP provides machine-speed negotiation across affected agents.

The Hive's cross-authority layer — peer coordination between sites, built on the same A2A foundation as the per-site Hive.

AGENT A Requesting 1. Request SCORECARD Impact Evaluation Full balanced scorecard 2. Authorize AGENT B AGENT C Affected agents 3. Consensus AUTHORIZED Execute Action Machine-speed: <100ms End-to-end negotiation

Multi-agent coordination pattern inspired by Google A2A Protocol and NVIDIA Multi-Agent Warehouse AI.

AWS Supply Chain Data Model

Entity compliance 35 / 35
ERP integrations SAP S/4, APO, ECC
Protocols RFC, OData, CSV
Proprietary formats None - fully portable

Escalation Arbiter

Inspired by dual-process cognitive theory, monitors agent decision quality over time.

DRIFT Detected ARBITER Evaluate OPERATIONAL Replan Parameters STRATEGIC Full Replanning Sustained drift (not one-off outliers) triggers routing

SCOR Metrics Hierarchy

Industry-standard SCOR (Supply Chain Operations Reference) metrics framework across three levels that cascade from strategic assessment to operational correction.

ASSESS - STRATEGIC Executive Health Check Revenue growth - EBIT margin - Return on capital Weekly / Monthly cadence DIAGNOSE - TACTICAL Cash-Flow Diagnostics Inventory turns - Cash-to-cash cycle - OTIF Daily cadence - Coordination layer CORRECT - OPERATIONAL Root Cause and Action Decision accuracy - Throughput - Override rate - Touchless rate Per-decision cadence - 11 execution agents Reliability Responsive Agility Cost Assets SCOR Attributes

Based on the ASCM SCOR Digital Standard. Metrics hierarchy follows the SCOR v12.0 Quick Reference.

Dive deeper

Talk to our team about the technical architecture behind Autonomy.