The Science of Deciding Without Waiting.

Each agent maintains a perception module — a continuously updated view of the portion of the operational environment relevant to its function. It subscribes to relevant data streams from the Omnierax event backbone, applies filtering and relevance scoring to distinguish signal from noise, and maintains a local belief state. Perception is not passive — agents can request additional data collection or clarification when their belief state uncertainty exceeds defined thresholds.

Agents maintain probabilistic belief states — not binary known/unknown representations, but probability distributions over possible states of the world, updated continuously as new evidence arrives. Bayesian inference is the foundational mechanism, extended with domain-specific priors derived from historical operational data and expert knowledge encoding. When evidence is contradictory, the belief state reflects the contradiction — agents do not silently discard inconvenient evidence. They flag it.

Each agent operates with a hierarchical goal structure — high-level mission objectives decomposed into sub-goals, further decomposed into specific action objectives. The goal structure is dynamic — it updates when the operational situation changes, when higher-level objectives are modified, or when an agent determines its current goal decomposition is no longer consistent with the achievable outcome space.

The agent's decision policy maps belief states and goal structures to action selections. Omnierax combines learned policies (from reinforcement learning on operational simulations and historical mission data) with structured decision logic (from domain expert knowledge encoding) — with the balance configurable based on criticality and novelty sensitivity. Novel situations are handled conservatively through structured logic; familiar patterns leverage learned efficiency.

Selected actions are passed to the action execution module — which validates that the action is within the agent's current authority scope, checks for conflicts with actions being taken by other agents, stages the action for execution, and monitors execution for success or failure. Failed actions trigger both local recovery sequences and escalation to the agent orchestration layer.

Operational intent — expressed in natural language or in structured policy format — is compiled into executable decision graphs by the Omnierax graph compiler. Compilation decomposes high-level objectives into specific decision nodes, identifies dependencies, defines information requirements, and generates the conditional logic that routes execution. Compiled graphs are validated against the current operational model before activation.

Four node types: Perception nodes collect and validate the information required for a decision. Inference nodes apply AI models to generate decision recommendations. Authorization nodes check whether the selected action is within current authority parameters — automatically escalating if not. Action nodes execute approved decisions and monitor outcomes, feeding results back into the graph.

Execution is event-driven — each node activates when its prerequisites complete and required inputs are available. Multiple branches execute in parallel where dependencies allow. The execution engine maintains complete state, enabling pause, resume, audit, and replay. When unexpected conditions appear, the engine applies pre-defined contingency logic or escalates to human review.

Decision graphs are not static once compiled. The adaptation engine continuously evaluates whether active executions are progressing as expected and whether the conditions present at compile time still hold. When significant changes are detected, it can modify the active graph — adding branches for emerging conditions, pruning irrelevant branches, or recommending recompilation for high-stakes operational shifts.

Runs continuously — every subscribed data stream is processed, every new event ingested and classified, every belief state updated. High-frequency sensor streams are processed in sub-100ms; periodic batch sources as they arrive; historical context maintained as a persistent background state that enriches real-time inference without reprocessing.

Above perception, the assessment cycle evaluates the current belief state against the active goal structure and identifies conditions warranting decision-making — opportunity, threat, or follow-up. Cycle rate is adaptive: faster in high-tempo environments, slower when the operational picture is stable, conserving compute while maintaining responsiveness.

When the assessment cycle triggers, the decision cycle generates candidate actions, evaluates them against the decision policy, selects the optimal candidate, and checks authorization. Output is either an authorized action (dispatched immediately) or a recommendation (routed to the HITL gateway). Latency from trigger to dispatch is a primary engineering metric.

Operates above the operational loop — continuously evaluating the quality of past decisions against their outcomes, identifying systematic biases or errors, and generating policy updates that are staged for review and deployment through the model governance framework. Autonomous performance improves continuously from operational experience — under human oversight, with updates validated before deployment.

For every decision, the explainability engine generates an evidence inventory — the complete list of data inputs that were considered, ranked by their contribution to the final output. Each item includes source, timestamp, confidence score, and the specific way in which it influenced the decision.

Documents the logical pathway from evidence to conclusion — the sequence of inferences, intermediate conclusions, alternative hypotheses considered, and the final determination process. Machine-generated but expressed in natural language operational users can read without specialized AI knowledge.

The confidence score is not a single global uncertainty estimate — it is a decomposed attribution that identifies which aspects of the decision are well-supported and which are uncertain. Decision-makers can calibrate their response based on where the uncertainty lies.

For significant decisions, the engine generates counterfactual analysis — identifying the minimal changes to input evidence that would have produced a different decision. Operators understand the sensitivity of the decision to specific inputs, and whether it is robust to reasonable variations in evidence quality or completeness.