How Agents Learn

The cognitive cycle

OODA, ORPA, and where Autonomy puts learning

A human operator running a facility moves through a loop on every consequential decision: observe the state, orient against context, decide, act, and start again. Boyd named this OODA. XMPro's MAGS architecture reformulates it for industrial AI agents as ORPA, Observe, Reflect, Plan, Act. The cycles are nearly the same: ORPA splits Boyd's Orient into a distinct Reflect step because in an agent architecture the sense-making computation deserves its own slot.

Autonomy's agents follow the same cognitive shape. Each agent observes canonical state, evaluates that state against its trained policy and the conformal calibration layer, plans an action within its envelope, and acts. The cycle is structurally identical to ORPA. The difference is not the shape of the cycle. It is where the learning happens.

The cycles are nearly identical. The learning placement is not.

OODA and ORPA both bake learning into the inner loop. Boyd's Orient step is where the observer's model gets updated; ORPA's Reflect step writes reflections into a memory stream that next-cycle reads. Each cycle does decision-making and learning in the same breath.

Autonomy lifts learning out of the cycle deliberately. The cognitive loop stays bounded, fast, and deterministic against the trained policy. Learning happens in a separate substrate: RL retraining runs on cadence against realised outcomes; conformal intervals recalibrate whenever measured coverage drifts; the causal layer updates override-effectiveness estimands as overrides accumulate; the LLM proposes hypothesis-axis CANDIDATEs that route to a human curator. None of those updates compete for the decision's compute budget; all of them are calibrated and outcome-supervised.

In-loop learning competes for compute with the decision itself and bakes the update into whatever the model's reasoning happened to be in that moment. Out-of-loop learning gets the cross-event time and aggregation it needs to be calibrated.

AI·IO·ML sits around the cycle, not inside it

AI·IO·ML, Autonomy's operating model, is not a replacement for OODA or ORPA. The agent still runs its own cognitive cycle on every decision. AI·IO·ML is the contract that governs the interaction between the substrate and the operator across three couplets: the agent acts (Automate, Inform), the human engages (Inspect, Override), and the system improves (Measure, Learn). The agent's cycle is its private business; AI·IO·ML is what the operator sees and acts on, and what the system measures and learns from. The same contract also governs the learning loop itself: parametric updates Automate when calibrated, Inform when they cross meaningful thresholds, and LLM-proposed CANDIDATEs always Inspect until the confidence head is calibrated.

Stage 1: Pre-Deployment, Twin Training

Deployment posture: Decision Support

Before an agent makes a single live decision, the platform has already built a digital twin of your supply chain DAG and used it to train the tier agents. A discrete event simulator runs your network forward under many plausible realities; deterministic engines modelled on ERP logic act as the teacher; the resulting (state, action, outcome) trajectories are what the tier agents learn from.

The execution agents do not need this customer-specific phase. They arrive warm-started from a generic corpus and walk their three-phase curriculum once live data is available.

Multiple realities, not Monte Carlo

Autonomy does not Monte-Carlo planning. Uncertainty around a plan is quantified by conformal P10/P50/P90 bands at inference time, not by re-running the twin under noise. The twin's role is to be a training environment, and its scenario sampler is curriculum-driven, not enumerated.

The curriculum shifts the mix over training, weighting heavier on baseline early (so the agent learns the dominant dynamics first) and tilting toward single-event and compound disruption later (so it learns to recover). Scenarios are stratified across four seasonal regimes, peak, ramp-down, trough, ramp-up, so the agent sees every season your business actually goes through.

Heuristics teach, they don't dictate

The deterministic engines that supervise Phase-1 training encode the same kind of rules your ERP already runs, base-stock-like ordering, lead-time-aware lot sizing, safety-stock allocation by percentage of on-hand, supplier-reliability-weighted re-order. But the agent never sees the rule. It sees the rule's output , what the heuristic chose to do in each scenario, and learns the underlying decision shape by example. This is the chess-engine pattern applied to inventory and supply: the rule sets the floor; the learned policy rises above it.

"Like a chess program that recognises common openings but makes illegal moves in novel positions, an agent trained on too little data will fail on unfamiliar situations. The twin's job is to push agents past that threshold before they ever touch live decisions."

By the time go-live happens, the tier agents have observed hundreds of thousands of simulated decision points across thousands of scenarios, not on a generic dataset, not on Monte Carlo enumeration, but on simulated rollouts of your network under disruptions calibrated to your history. Day-one performance matches or exceeds the heuristic baseline that runs in your ERP today.

Stage 3: Continuous, CDC-Driven Retraining

Deployment posture: Decision Automation

Operations change. Suppliers shift lead times, demand patterns evolve seasonally, new products arrive, old products retire. An agent trained on last quarter's dynamics will drift if it is not maintained. Autonomy detects drift and retrains automatically.

Baseline cadences

The platform retrains on a fixed schedule regardless of whether drift is detected. Conformal coverage drift detection runs independently of that schedule and can accelerate any of these cadences when it triggers.

These are defaults. Every tenant has its own training-cadence configuration, so a distributor with high-velocity demand can pull execution-agent and L3-model retraining in to 6-hourly; a pharma manufacturer with stable, regulated planning can push S&OP out further. The cadence knob is per-agent-type and per-tenant, not a platform constant.

Regression guard

Retraining is not deployed blindly. Every candidate checkpoint is evaluated against the current production model on held-out trajectories. If the new model regresses on BSC utility, it is discarded; only checkpoints that match or improve get promoted.

Conformal prediction, never Monte Carlo

Uncertainty around any plan is quantified by conformal P10/P50/P90 bands at inference time. The digital twin is never re-run under noise to "Monte Carlo a plan". This is enforced as an invariant in the platform, the twin's PLAN_PRODUCTION mode raises an error if any stochasticity knob is left on. Drift detection then watches the empirical coverage of those bands: if the predictor promised 90% coverage and the observed coverage falls materially below it, the bands are recalibrated and the upstream policy is retrained.

Escalation

If retraining at one tier cannot close the drift, a structural change like a supplier permanently doubling lead times, the failure escalates upward. The L1 execution agent cannot fix policy; the tactical (L3) Domain-Model Reconciliation model re-optimises buffer levels; the strategic (L4) Policy Optimisation model re-evaluates network design. The escalation chain runs automatically.

The compounding advantage

Two things compound as a customer runs longer on Autonomy. The first is the tier agents' familiarity with your DAG, every season they see, every disruption they recover from, every supplier behaviour they internalise makes the next decision better. The second is the Operating Knowledge layer built from your planners' overrides and the outcomes that followed.

A competitor adopting the same platform would arrive at Phase 1: their execution agents warm-started from the same generic corpus, their tier agents trained on their DAG and seasons. Their agents would be competent on day one, just as yours were. But they would not have your years of seasonal cycles in their L3 model weights, and they would not have your planners' decade of override patterns shaping their execution agents. The platform is the same; the trained-on-your-business specificity is not.