Azirella
← Back to Autonomy

Architecture

Context & Decision Architecture

Autonomy is not a single model making decisions. It is a hierarchy of specialized agents, each operating at a different time scale, where context, guardrails and targets cascade downward from strategic to execution, and feedback flows upward from the front line to reshape strategy. Everything reads from and writes to one shared world model, and judgment calls surface through the single Decision Stream under AIIO.

"The most significant shift in enterprise AI is not from manual to automated, but from tool-assisted to agent-directed. Organizations that treat AI as a copilot will be outcompeted by those that architect for autonomous execution with human governance."

, Jordi Visser, "The Agentic Inversion"

The Decision Hierarchy

Five tiers, each with its own cadence, scope, and authority

CONTEXT ENGINE Documents · Natural Language Directives · Email Signals · Market Data Parses, classifies, routes to the appropriate tier Strategic: Strategic, Network Planning Weekly/Monthly · Network-wide · Policy Optimization Outputs: Safety stock multipliers · Service level targets · Reorder points Sourcing splits · Allocation priorities · Network risk profiles Sets the guardrails and KPI targets that constrain every tier below Policy parameters + network risk profiles Tactical: Network Coordination, Demand / Supply / Inventory Daily · Cross-site · Demand/Supply/Inventory Planning Outputs: Site-specific directives · Demand forecasts · Exception probabilities Allocation adjustments · Propagation impact scores · Order recommendations Translates network-wide strategy into site-specific goals and constraints Site directive per site Operational: Site Coordination (per site) Hourly · Per site · Cross-agent coordination Outputs: Urgency adjustments per execution agent · Priority conflict resolution Resource arbitration across competing agents within a single site Ensures 11 execution agents within a site don't work at cross-purposes Urgency vectors + guardrails + targets Execution: Execution, 11 Agents (per site) <10ms · Per decision · Pattern-matched AATP · PO · MO · TO · Quality · Maintenance · Rebalancing · Forecast Adj Subcontracting · Inventory Buffer · Order Tracking Each agent decides within guardrails set by the tiers above MPS/MRP inputs Deterministic Planning Engine (off-hierarchy) MPS/MRP · BOM Explosion · Net Requirements · Safety stock · Lot sizing Generates training data for the four tiers. Not in the live decision loop. CONTEXT & GUARDRAILS DOWN FEEDBACK & OUTCOMES UP

The Context Engine

The Context Engine is the system's sensory cortex. It ingests signals from three channels and routes them to the appropriate tier of the decision hierarchy:

Documents

Strategy reports, policies, planning guidelines uploaded by leadership. Parsed and converted into structured parameters that feed strategic policy optimization.

Azirella Assistant

Natural language directives from planners and executives. "Increase safety stock for electronics by 20% for Q4" is parsed, validated, and routed to the right decision tier based on the speaker's role.

Email Signals

Automated signal intelligence from supplier and customer emails. Demand changes, supply disruptions, lead time shifts, quality issues, classified and injected into the signal bus. GDPR-safe: all PII stripped before persistence.

The Context Engine doesn't just collect data, it routes it. An executive directive about service level targets routes to Strategic. A supplier email about a delayed shipment routes to the site's execution agents. A planner's instruction to prioritize a key customer routes to the tactical network planners. The routing is role-aware: VPs and executives influence network-wide policy; analysts influence individual execution decisions.

Strategic, Strategic: Setting the Guardrails

The strategic planning agent operates on a weekly cadence across the entire network. It analyzes the supply chain topology, which sites are critical, where bottleneck risk concentrates, how resilient each node is, and produces policy parameters that constrain every tier below:

Safety Stock Multipliers

How much buffer each site should carry, calibrated to its criticality in the network.

Service Level Targets

Fill rate and OTIF goals per site, balancing cost against customer commitments.

Reorder Points & Order-Up-To Levels

When to trigger replenishment and how much to order, the guardrails for purchasing agents.

Sourcing Splits & Allocation Priorities

Make-vs-buy ratios and customer tier priorities that shape how inventory is allocated.

These parameters are optimized across thousands of Monte Carlo scenarios — exploring a wide range of possible demand/supply futures to find the policy that performs best on average. The output includes a network risk profile per site that encodes the topology, risks, and priorities. This profile flows down to every tier below as context.

When an executive says "optimize for service level this quarter," the Context Engine routes that directive to Strategic, which adjusts the reward weights in its optimization, shifting the trade-off from cost toward fill rate. The updated policy parameters cascade downward, and within a day every agent in the network is operating under the new priorities. No manual reconfiguration required.

Tactical, Network Coordination: Translating Strategy to Sites

The tactical planning agents (Demand, Supply, Inventory) run daily. They consume the strategic policy parameters and translate network-wide strategy into site-specific directives:

Site Directive, what each site receives daily
Demand forecast, 4-period ahead, specific to this site's products
Exception probability, stockout, overstock, and normal likelihood
Propagation impact, if a disruption hits this site, how far does it cascade?
Order recommendation, suggested replenishment quantity
Allocation adjustments, customer priority rebalancing for this site
Inter-hive signals, coordination signals from neighboring sites

Tactical is where network-wide intelligence meets local reality. Strategic might set a 95% service level target for a distribution center, but Tactical knows that this DC's primary supplier has been unreliable for the past two weeks. It adjusts the site directive accordingly, increasing the exception probability score, flagging propagation risk to downstream sites, and recommending a larger order quantity to compensate for supply uncertainty.

This is commander's intent, not micromanagement. Tactical tells each site what to achieve and what to watch for, not how to do it. Tiers 3 and 2 retain full authority over how to meet the goals.

Operational, Site Coordination (per site)

Each site has a site coordinator agent that runs hourly. Its job is coordination: ensuring that the 11 execution agents within a site don't work at cross-purposes.

Consider a site where the ATP agent is committing inventory to new orders while the Rebalancing agent is trying to transfer that same inventory to a neighboring DC. Or where the PO agent is ordering more material while the Inventory Buffer agent is trying to reduce stock. These conflicts are natural, each agent optimizes for its own objective. The site coordinator resolves them.

It does this through urgency adjustments, a vector of 11 values (one per execution agent) that modulates each agent's priority. When production capacity is constrained, the MO Execution agent's urgency increases while the PO agent's decreases. When a key customer has an urgent order, the ATP agent's urgency spikes. The 5-tier priority system ensures critical functions (customer commitments, production safety) always take precedence:

  1. 1 Customer commitments, ATP, Order Tracking
  2. 2 Production safety, MO Execution, Quality
  3. 3 Supply continuity, PO Creation, TO Execution
  4. 4 Lateral flow, Rebalancing, Subcontracting
  5. 5 Planning refinement, Forecast Adjustment, Inventory Buffer, Maintenance

Execution, Execution: Decisions in Under 10ms

The 11 execution role agents at each site make the actual decisions, allocating ATP, creating purchase orders, scheduling manufacturing, routing transfer orders, managing quality holds, and more. Each decision takes under 10 milliseconds.

But these agents don't decide in isolation. Every decision is shaped by the full context cascade:

Example: PO Creation Agent deciding order quantity
From Strategic: Reorder point = 14 days, order-up-to = 28 days, safety stock multiplier = 1.3x
From Tactical: Demand forecast shows 15% increase next period, propagation impact = high (hub site)
From Operational: Urgency elevated (inventory trending below buffer), PO priority boosted above rebalancing
From Signal Bus: Neighboring site emitted LATERAL_SHORTAGE signal; Order Tracking flagged a late inbound shipment
Decision: Order 1,850 units (vs. baseline 1,200) with expedited delivery. Confidence: 0.82. Urgency: 0.71.

The agents communicate laterally through the Hive Signal Bus, a pheromone-inspired messaging system where signals decay over time (30-minute half-life). When the ATP agent detects a shortage, it emits an ATP_SHORTAGE signal. The PO agent reads this signal and factors it into its next order decision. The Rebalancing agent reads it and considers pulling inventory from a neighboring site. All of this happens within milliseconds, without any central coordinator.

The Feedback Loop, Intelligence Flowing Upward

Context flows down. Feedback flows up. Every execution decision generates outcome data that feeds back to reshape the tiers above. This is what makes the system adaptive rather than merely hierarchical.

What Execution Agents Report Upward

Each site's agent hive produces a structured feedback report that the tactical and strategic tiers consume:

Performance Metrics

Fill rate, OTIF, backlog change, inventory position, how well the site is performing against targets.

Decision Confidence

Average confidence across all agents, override rate, how often agents disagreed with human planners.

Signal Activity

Shortage vs. relief signal counts, net urgency trend, dominant signal types, the "mood" of the site.

CDC Triggers

How many data drift thresholds were breached, maximum severity, early warning of structural change.

A Concrete Example

The MO Execution agent at Plant B has been running at 95% OEE for the past 3 weeks. It flags this as unsustainable, equipment fatigue, deferred maintenance, and increasing quality holds are accumulating. The signal propagates upward:

1
Execution agent flags
MO agent emits MAINTENANCE_URGENT signal with high urgency. Quality agent confirms with QUALITY_HOLD signals trending upward.
2
Operational responds
Site coordinator increases Maintenance agent's urgency, decreases MO agent's production targets to allow maintenance window. Other agents at this site adjust to the reduced capacity.
3
Tactical adapts
Tactical planning agents detect reduced capacity at Plant B. Redirect some production load to Plant A and DC East via cross-site coordination. Adjust demand fulfillment routing across the network.
4
Strategic learns
If the pattern persists, escalation triggers strategic re-optimization. Policy parameters are updated, perhaps Plant B's capacity assumption is permanently reduced, safety stock at downstream sites increased.

This entire cascade, from a frontline agent noticing an unsustainable trend to network-wide policy adjustment, happens automatically. No meetings, no manual escalation, no waiting for the next S&OP cycle.

Escalation, When Lower Tiers Can't Fix It

The system follows a dual-process model. Execution agents are the fast track — pattern-matched, handling 95% of decisions automatically. The higher tiers are the slow track, analytical, activated only when the fast track fails.

The Escalation Arbiter runs every 2 hours, watching for persistence: when execution agents have been consistently adjusting in the same direction for 48+ hours, it signals that the current policy parameters are wrong — the world has shifted beyond what execution-level retraining can fix.

Escalation Routing
Single agent, <24h: Normal noise, no action, CDC handles it
Single agent, >48h: Local policy drift, operational site refresh
Multiple agents, same site: Site-level issue, operational refresh + rebalance
2-3 sites, same direction: Regional shift, tactical planning refresh
>30% of sites, same direction: Network-wide shift, S&OP strategic review

The key insight: authority is pushed to the lowest capable level. The Escalation Arbiter never overrides a lower tier, it requests replanning at the appropriate higher tier. Like a military chain of command: frontline soldiers (execution agents) have full autonomy within their rules of engagement (guardrails). When the situation exceeds their authority, it escalates, not to override them, but to give them better rules of engagement.

Thinking Like an Organization

The architecture draws from two complementary frameworks:

Dual-Process Decision Making

System 1 (execution agents): Fast, intuitive, pattern-matched. Handles routine decisions in <10ms. Only escalates when it can't cope.

System 2 (Tiers 4-5): Slow, analytical, deliberate. Activates on anomalies, processes cross-site patterns, re-optimizes policies.

Nested Observe-Orient-Decide-Act Loops

Execution loop: <10ms. Observe local state → Orient via urgency vectors → Decide → Act immediately.

Tactical loop: Daily. Observe cross-site patterns → Orient via demand/supply models → Decide allocations → Push directives.

Strategic loop: Weekly. Observe network topology → Orient via risk analysis → Decide policy parameters → Propagate to all tiers.

The result is a system that operates at human-impossible tempo at the edges (thousands of decisions per second across all sites) while maintaining strategic coherence at the center (weekly policy optimization informed by months of accumulated feedback). Context cascades down. Feedback flows up. The organization thinks.

See the Architecture in Action

Watch context cascade through five tiers in a live supply chain demo.