Named agent identity
Distinct & addressableEvery agent is a distinct, addressable identity, not an anonymous background process. You can always see which agent reasoned, recommended, or acted.
HOW XMPRO MAGS WORKS
The decision architecture behind every agent.
MAGS is a governed decision architecture, not a single model. Here are the parts that make it work.
The cognitive decision loop Observe, Reflect, Plan, Act. Agents decide, they don’t just respond. Cognitive intelligence Memory, confidence, and reasoning: how agents make sense of what they see. How the team coordinates Each agent runs its own loop; the team agrees the decision before it acts. Planning & optimisation Plans built against objective functions and constraints. Optimising, not reacting. Governed action Actions route through guardrails, simulation, and human authority. Bounded by design. The Decision Trace Every decision leaves evidence: what was seen, why it mattered, what governed it.
THE COGNITIVE DECISION LOOP
XMPro's Unique Cognitive Approach
to Multi-Agent Generative Systems
XMPro doesn't just add AI to operations — it re-architects how decisions are made, shared, and executed across the enterprise.
RELIABLE DECISION MAKING
Observe. Reflect. Plan. Act.
PLATFORM FIT
MAGS sits between trusted context and governed action.
SignalsData Stream Designer
Trusted contextOCE
Decision teamXMPro MAGS
Governed actionFRS · human review · Action Agents
EvidenceDecision Trace
The architecture behind it
Decision-making stays separate from execution, so industrial AI acts with intelligence, not risk.
COGNITIVE INTELLIGENCE
Making sense of what the agent sees.
Cognitive intelligence is how a MAGS agent perceives, remembers, and reasons: it turns raw signals into operating knowledge it can act on, not just a prompt it responds to.
Memory significance
Scores which observations matter, by recency, importance, and relevance.
Memory & retrieval
A persistent operating-history stream, retrieved by relevance, not just recency.
Synthetic memory
Distils raw observations into higher-order memory the agent can reason over.
Confidence scoring
How sure the agent is, computed from the evidence, not asserted by the model.
Content processing & strategy
Interprets and structures incoming content into usable operating context.
Plan-adaptation detection
Detects when conditions have shifted enough that a plan needs to change.
HOW THE TEAM COORDINATES
Each agent runs its own loop. The team coordinates the decision.
In a MAGS decision team, each specialist agent runs its own ORPA loop. A safety agent does not think like an economic agent; a reliability agent does not evaluate a situation the same way as a process-optimisation agent. That separation prevents one generic agent from holding every domain, objective, and constraint in a single prompt.
Governed coordination
Agents publish observations, plans, and intent to a governed message broker, by subscription rather than direct calls. No uncontrolled back-and-forth.
Consensus before action
When a decision needs team alignment, agents run collaborative iteration: conflict detection, multi-round resolution, and a formal vote before anything acts.
Lifecycle & governance
Each agent’s role, state, and governance are managed across the team, so coordination never dissolves individual accountability.
PLANNING & OPTIMISATION
Plans are optimised, not improvised.
A MAGS agent doesn’t react to the latest reading. It scores each candidate action against an objective and a utility function, inside the limits its constraints allow, and optimises for the whole operation over time rather than the next alarm.
Competing objectives (safety, quality, throughput, energy, cost) are weighed by explicit priority, with safety ranked ahead of throughput. The constraints define what is feasible before optimisation even begins.
Objective functions
Define what the agent or team optimises: a business goal written as a scored target it reasons against, not a hidden prompt.
Plan optimisation
Searches and scores candidate plans against the objectives and constraints, choosing the action that serves the whole operation.
Performance monitoring
Measures the outcome of each decision and tracks performance over time, so the objective stays tuned to reality.
GOVERNED ACTION
The model proposes. The policy disposes.
An agent can plan an action without being allowed to take it. Every action leaves through a configured, governed pathway, or it does not leave at all, and stays attributable to the agent that took it and the person accountable for it.
Bounded autonomy
Hard limits are obligatory, forbidden, and permitted rules, checked deterministically before anything acts. The agent’s learning can never move the safety envelope.
Governed pathways
Every action routes through a configured pathway: human review, approval, FRS simulation, an Action Agent, or escalation. No pathway, no action.
Evidence & escalation
A blocked action escalates or requests approval, and the reason is recorded. Nothing is silently dropped.
THE DECISION TRACE
The decision loop leaves evidence.
Decision Trace records what the agent observed, why it reflected, what it planned, what action intent was routed, and which constraints applied. Industrial agentic systems need reviewability: operations need to know what happened, engineering needs the decision path, and governance needs evidence the agent stayed inside its boundaries. It also supports learning — which reflections created value, which plans were blocked, and where the operating envelope should be refined.
Every trace records
- Observation
- Reflection
- Plan
- Action or action intent
- Timing
- Confidence and significance
- Tool activity
- Constraints and policy checks
- Human review or approval
- Agent role and context
- Outcome
WHAT THIS MEANS FOR OPERATIONS
A governed path from monitoring to action.
Many industrial AI projects stop at insight. They tell a human what may be happening, then leave the hard work of context, coordination, approval, and action to the existing organisation. MAGS is designed to close more of that gap without removing governance.
The same decision loop can support recommendations, human-approved actions, and policy-controlled bounded autonomy as the use case matures. The operating benefit is consistency: the team watches continuously, evaluates repeat decisions against the same objective and constraint logic, and records the evidence every time.
AGENT IDENTITY & ACCOUNTABILITY
Every agent has an identity, and an owner.
As agents multiply across an operation, most have no durable identity, so no one can say which agent acted or on whose authority. XMPro treats agent identity as a first-class governance control.
A named human owner
Delegated authorityEach agent is tethered to an accountable person or role. Authority is delegated deliberately, never left orphaned.
Enterprise identity, federated
Entra ID readyAgent identities integrate with enterprise directory and access systems such as Microsoft Entra ID, so agents are provisioned, permissioned, and revoked like any other identity.
Attributable by default
Into the Decision TraceEvery action carries its acting agent and accountable owner into the Decision Trace, so nothing an agent does is left unattributable.
You cannot govern, audit, or trust an agent you cannot name, or hold to account through the person who owns it.
You’ve seen how MAGS works.
See it running in a real deployment, or talk to our team about the decisions you’d put it to work on.