The platform
The Autonomy Platform
One Decision Stream. Demand to delivery. One shared world model under all of it.
Autonomy is a Decision Intelligence platform for the full supply chain, six decision domains, executed by coordinated agents, governed by AI·IO·ML, and surfaced through a single Decision Stream with urgency × likelihood × triage.
Latency Compounds Costs. Decision Velocity Compounds Value.
Every layer of waiting between signal and action compounds cost on top of cost. Every agent at every tier, from strategic network design to sub-second execution, cycles through the same OODA loop: Observe, Orient, Decide, Act. The faster each agent cycles, the greater your competitive advantage. Outcomes feed back as a learning signal, making every cycle smarter than the last.
Every agent, every tier, every cycle
The agent reads its current state, inventory levels, demand signals, capacity, upstream decisions, and context from the tiers above.
The agent evaluates alternatives, runs its learned policy, assesses trade-offs, and calibrates likelihood using conformal prediction.
The agent commits to its best action, or surfaces it in the Decision Stream when likelihood is low and urgency is high.
The agent executes. The outcome is measured against the balanced scorecard and fed back as a learning signal.
Strategic agents cycle weekly. Tactical agents cycle daily. Operational agents cycle hourly. Execution agents cycle in <10ms. All four tiers learn continuously from outcomes, and human overrides and coaching accelerate the improvement.
John Boyd's insight: The competitor who cycles through OODA faster doesn't just react quicker, they shape the environment. Autonomy compresses the decision loop from days to seconds, turning your supply chain from reactive to anticipatory.
Colonel John Boyd, "Patterns of Conflict" (1986)
"Planning systems have never had a complete knowledge model of the supply chain they were supposed to plan. The planner was the missing ontological layer, semantic intelligence turning data into understanding."
Autonomy supplies that missing layer through the shared world model, one ontological graph every agent reads from and writes to.
Four-Tier Decision Architecture
Each tier handles decisions at its natural time scale. Context and guardrails cascade downward; escalation and feedback flow upward, a continuous learning loop.
Strategic / Network
Weekly/monthly, network-wide. Policy parameters, risk scoring, portfolio optimization, KPI targets, and bottleneck detection.
Tactical / Network
Daily, cross-site. Demand & supply balancing, priority allocations, cross-site coordination, and context for execution agents.
Operational / Site Coordination
Hourly, per site. Cross-agent coordination, cascade prediction, urgency adjustments, and cross-functional trade-offs.
Execution / Site & Role
<10ms, per site, 11 agents. Narrow decisions: ATP, purchase orders, rebalancing, manufacturing, quality, maintenance.
How the four pillars compose
Any vendor can list these four pillars: AI agents, conformal prediction, digital twin, causal AI. Each is well understood and replicable. The substrate is the composition discipline that makes them produce one decision system, not four siloed libraries.
Twin × Causal
Calibrated priors on day one
The digital twin generates synthetic supply-chain scenarios. Every learned component in the causal layer pre-trains against them. A new decision class lands with a calibrated prior on day one instead of waiting 6-18 months for real overrides to accumulate.
Causal × Agent
Overrides become a learning signal
When a planner overrides an agent, the causal layer estimates the treatment effect against the agent's counterfactual. The next retrain cycle up-weights overrides from planners whose interventions actually improve outcomes. Tacit knowledge becomes substrate knowledge.
Agent × Conformal
Every decision carries an honest band
Every TRM output is a P10 / Median / P90 distribution, not a scalar. The AI·IO·ML Inform threshold reads the calibrated band's width to decide whether to act, surface, or hold. Operators see when the agent is confident and when it is not.
Three more pairwise compositions
Twin × Agent (generic pre-training of the decision tier), Conformal × Causal (valid bands on every causal claim regardless of model misspecification), and Twin × Conformal (the conformal coverage promise holds from day one because the twin supplied the calibration set).
Two whole-stack compositions
Training-time: one nightly digital-twin rollout fans output to all four pillars in parallel, producing one calibrated stack instead of four siloed artefacts. Decision-time: every Decision Trace row carries the four-field AI·IO·ML Inspect contract, with each pillar contributing one field. Every row is a four-pillar joint output, never four signals next to each other.
The substrate is the discipline, not the pillars. Every learned component co-trains on the same twin rollouts. Every output band conformally calibrates. Every decision row populates all four Inspect-contract fields. Every outcome causally attributes back to a decision. A competitor with the same four pillars but no composition discipline produces a stack that looks similar and behaves differently. Customers experience the difference at audit time, at onboarding time, and at the moment they try to add the next decision class.
Capabilities
The building blocks under the Decision Stream.
Decision Intelligence
Full Gartner DI lifecycle, model, orchestrate, monitor and govern every supply chain decision as a first-class asset in the Decision Stream.
Demand & Supply Planning
Statistical and ML forecasting, net requirements, multi-level BOM explosion, and multi-sourcing with priorities, running against the shared world model.
AI Agents
11 specialized agents execute narrow decisions at machine speed. Four-tier architecture from strategic network design to site-level execution.
Execution
Available-to-Promise, order management, manufacturing and transfer orders, quality disposition, and maintenance scheduling, all in one stream.
Stochastic Planning
20 distribution types and Monte Carlo simulation generate the scenario data for agent training and the calibration sets for conformal prediction.
Conformal Prediction
Distribution-free likelihood guarantees for every agent decision. Calibrated coverage that holds regardless of data distribution or model quality.
Causal AI
Counterfactual reasoning determines whether decisions actually caused outcomes. The only rigorous way to separate skill from luck and drive genuine improvement.
Digital Twin Simulation
Simulation and validation environment for training agents and building team confidence.
Context Engine
Five input channels, internal documents, external data, communication signals, voice/text directives, and IoT events, ground every agent decision in organizational context.
Operating Knowledge
Captures the tacit behavioral knowledge experienced planners carry. Overrides are classified as GENUINE or COMPENSATING and fed into agent training, preserving institutional expertise.