REF-01 / System Architecture

CORTEX
Architecture

Hub-and-spoke multi-agent topology with strict domain separation. 14 agents, one integration layer, zero redundancy by design.

Architecture Manifest
cortex.arch / v1.0
01patternhub_and_spoke
02agents.total14
03agents.tier_1_core5 // primary manufacturing domains
04agents.tier_2_domain6 // facility-specific domains
05agents.tier_3_support3 // monitoring, routing, escalation
06domain_separationstrict // one domain per agent, no exceptions
07inter_agent_commsasync_message_passing
08integration.primaryMicrosoft Teams
09integration.dataCustom spreadsheet schemas per stakeholder
10escalation.modelthreshold_based // SENTINEL owns escalation
11data_gravity.window90 days
12build_time~6 hours // concept to operational

Hub-and-Spoke Topology

CORTEX uses a hub-and-spoke topology. A central coordination layer receives inputs from all 14 agents and distributes directives downward. Agents do not communicate with each other directly — all inter-agent coordination passes through the hub.

This design decision was deliberate: direct peer-to-peer agent communication creates emergent coupling that makes the system harder to debug, audit, and remove. Forcing all coordination through a central point also makes the system's behavior transparent to stakeholders who interact with the hub interface.

Design principle: An agent that communicates with two other agents directly is two coupling points that will eventually produce unexpected behavior. All coordination through the hub. No exceptions.

Topology Specification

Hub layer
Central coordination point. Receives agent reports, resolves conflicts between domains, distributes context updates, and owns the external stakeholder interface.
Spoke agents
All 14 CORTEX agents. Each agent has one connection: upward to the hub. No lateral connections. Each agent owns one domain and reports one class of information.
Report cadence
Agents report on event-trigger or scheduled interval depending on domain type. YIELD reports on production events. WATT reports on scheduled energy intervals. SENTINEL reports on threshold crossing.
Directive flow
Directives flow downward from hub to agents. Agents do not issue directives to each other. An agent can request hub intervention by escalating, which may result in a directive to another agent via the hub.

Domain Separation

Domain separation is the central architectural constraint. Each agent owns exactly one domain. If an agent begins performing two functions, that is a design failure — those functions belong to two separate agents that have not been separated yet.

This constraint was enforced during the initial 6-hour build and is not negotiable in any subsequent configuration change. The naming convention reinforces this: YIELD owns yield. WATT owns energy. WRENCH owns maintenance. The name is the domain. The domain is the scope. The scope does not expand.

Domain Separation Rules
cortex.constraints / v1.0
01ruleONE_DOMAIN_PER_AGENT
02enforcementstrict
03violation_responseDecompose agent into separate agents
04scope_expansionPROHIBITED
05cross_domain_queriesRoute through hub, not direct agent-to-agent
06naming_conventionAgent name == domain name == scope boundary

Communication Protocol

Agents communicate with the hub via asynchronous message passing. Each message has a type, a source agent, a priority level, and a payload. The hub processes messages from a priority queue: SENTINEL escalations take precedence over routine domain reports.

Message Types

REPORT
Routine domain status update from agent to hub. Scheduled or event-triggered. Priority: normal. Hub acknowledges and updates state model.
ESCALATION
Threshold-crossing event requiring hub attention. Issued by SENTINEL or any agent detecting an out-of-bounds condition. Priority: high. Hub routes to appropriate stakeholder interface.
DIRECTIVE
Hub-to-agent instruction. May be a scheduled task assignment, a context update, or a response to an escalation. Agents execute directives without negotiation.
QUERY
Cross-domain information request. Routed through hub. Hub retrieves from the target agent's last REPORT and returns without requiring direct agent-to-agent contact.

Integration Layer

CORTEX integrates with the EP44 operational environment through Microsoft Teams. This integration decision was made deliberately: agents live where the work lives. Stakeholders do not adopt new tools — CORTEX appears inside the tool they already use.

Each stakeholder has a personal agent configured to their workflow. The agent learns the spreadsheet schemas, report formats, and terminology specific to that person's role. Over time, the agent becomes a collaborator inside the existing workflow rather than an external system requiring login and context switching.

Data gravity mechanism: By learning custom schemas per stakeholder and embedding in Teams, CORTEX accumulates irreplaceable institutional context. After 90 days, that context is deeply embedded in the facility workflow. Removal becomes operationally disruptive before it becomes politically difficult.
Integration Specification
cortex.integration / v1.0
01platformMicrosoft Teams
02agent_presencePersonal agent per stakeholder
03schema_learningactive // per-user spreadsheet patterns
04gravity_window90 days
05context_switchingnone required // embedded in existing workflow
06stakeholder_onboardingPassive — agent adapts to user, not inverse
Next: Agent Catalog → ← Back to Index