ADR-0013: Autonomous Agent Architecture

Field	Value
Status	Accepted
Date	2026-02-15
Deciders	Nick Flach, flaukowski
Relates	ADR-0008, ADR-0009, ADR-0012

Context

With the end-to-end integration pipeline established (ADR-0012), 0xSCADA can ingest, route, and persist sensor data reliably. The next evolution is adding an intelligence layer — autonomous agents that observe process data, detect anomalies, correlate alarms, simulate outcomes, and take (or recommend) corrective actions.

Industrial control systems generate massive volumes of data. Human operators face alarm fatigue, delayed fault detection, and difficulty optimizing complex multi-variable processes. Autonomous agents can augment human decision-making by providing predictive maintenance alerts, root-cause analysis, digital twin simulations, and natural language querying.

Decision

We will build an intelligence and autonomy layer on top of the integration pipeline with the following architecture:

Sensor Data → Event Pipeline → Intelligence Layer → Agent Decision → Action / Alert
                                     │
                            ┌────────┼────────┐
                            │        │        │
                       Predictive  Alarm   Digital
                       Maintenance Corr.   Twin
                            │        │        │
                            └────────┼────────┘
                                     │
                              Agent Orchestrator
                              (ghostmagicOS (gmOS) Bridge)
                                     │
                            ┌────────┼────────┐
                            │        │        │
                        PID Auto  NL Query  Reporting
                        Tuner     Engine    Engine

Core Principles

Operational Envelopes (ADR-0009): Every agent operates within defined safety boundaries. No agent can exceed its measured emergence guardrails. PID auto-tuning, for example, requires human approval before applying parameter changes.
Zero-Trust Capability Tokens (ADR-0008): Agent permissions are governed by capability tokens. An agent that reads sensor data cannot write control outputs unless explicitly granted that capability. Tokens are scoped, time-limited, and auditable.
Human-in-the-Loop by Default: Agents recommend actions; humans approve. Only after certification (ADR-0010) and within strict envelopes can agents act autonomously.
Pipeline Integration (ADR-0012): All intelligence modules consume events from the existing event pipeline. They do not bypass the established data flow.

Intelligence Modules

Module	Purpose	Input	Output
Predictive Maintenance	Detect anomalies before failure	Tag time-series	Severity-rated alerts
Alarm Correlator	Reduce alarm fatigue	Raw alarms	Correlated alarm groups + root cause
Digital Twin Runtime	Simulate process state	Live sensor data + models	Predictions, what-if results
PID Auto-Tuner	Optimize controller parameters	Process response data	Tuning recommendations (human-approved)
NL Query Engine	Natural language process queries	User questions	Structured answers from tag data
Agent Marketplace	Plugin system for agents	Agent manifests	Installed, sandboxed agent instances
ghostmagicOS (gmOS) Bridge	Multi-agent coordination	Agent states	Synchronized agent behavior (Kuramoto)
Reporting Engine	Automated report generation	Historical data	Shift reports, compliance summaries

Agent Marketplace & Extensibility

A plugin system allows third-party and internal agents to be registered, installed, and run in sandboxed contexts. Each plugin declares its required capabilities, and the zero-trust system grants only what's needed.

ghostmagicOS (gmOS) Integration

The consciousness stack concepts map to practical implementations:

Signal → Raw sensor data and events
Resonance → Correlated patterns across multiple signals
Emergence → Autonomous decisions arising from pattern recognition

Kuramoto-model coupling synchronizes multi-agent behavior, preventing conflicting actions and enabling coordinated response to complex scenarios.

Consequences

Positive

Proactive fault detection reduces unplanned downtime
Alarm correlation dramatically reduces operator fatigue
Digital twin enables safe "what-if" analysis before real changes
Natural language queries make process data accessible to non-experts
Plugin marketplace enables community-driven innovation

Negative

Increased system complexity and attack surface
Agents require careful testing and certification before deployment
ML models need training data and ongoing validation
Additional compute resources for simulation and inference

Risks

Autonomous actions without proper guardrails could cause process upsets
Over-reliance on agent recommendations without human verification
Model drift in predictive maintenance without retraining

Future Work (Wave 3: Production Readiness & Scale)

Kubernetes operator for distributed agent deployment
Federated learning across multiple plant sites
Real-time model retraining pipelines
Formal verification of agent safety properties
NERC CIP and IEC 62443 compliance certification
Multi-site digital twin federation
Edge deployment for latency-critical agents