Stack: .NET 10 · ASP.NET Core · Next.js 15 · Azure · Kubernetes
Enterprise-grade multi-layer AI platform — structured reasoning, multi-agent orchestration, zero-trust security, knowledge management (RAG), plugin-based UI, and full NIST RMF compliance in a single .NET 10 + Next.js solution.
cognitive-mesh is the intelligence and governance core of the phoenixvc platform. It is a complete enterprise AI platform structured as seven interdependent layers, each with its own domain logic, ports/adapters, and contributing agents. Think of it as the full-stack operating system for intelligent, governed AI workloads — from raw infrastructure (data stores, secrets, audit logs) up to a user-facing plugin dashboard.
It is the "thinking layer" that sits between raw model calls (via sluice) and the products that need intelligent, multi-step results. Every other phoenixvc service either feeds it data, routes work through it, or consumes its outputs.
Everything the upper layers depend on, all behind ports so implementations are swappable.
- Zero-Trust Security: JWT auth, RBAC/ABAC policy enforcement, secrets management (Azure Key Vault), audit logging
- Data persistence: Azure Cosmos DB adapter, Azure Blob Storage adapter, Vector Database abstraction (Azure AI Search / Redis)
- Knowledge & RAG: Semantic search via
EnhancedRAGSystem, Knowledge Graph, Document Processing / ingestion pipeline - Enterprise integration: Microsoft Fabric OneLake adapter, Feature Flag manager, Enterprise Connectors
Modular reasoning engines, each exposed via a typed port interface. Higher layers call these engines without knowing their implementation.
- StructuredReasoning (ConclAIve): Converts raw LLM outputs into auditable structured reasoning via Debate & Vote, Sequential Reasoning, and Strategic Simulation recipes
- AnalyticalReasoning: Data-driven analysis, trend identification, structured insight generation
- SecurityReasoning: Threat intelligence engine, anomaly detection, dynamic risk scoring
- EthicalReasoning: Normative agency (Brandom), information ethics (Floridi) — validates actions against ethical frameworks
- CreativeReasoning / CriticalReasoning: Idea generation, logical consistency evaluation, bias detection
- DomainSpecificReasoning: Pluggable industry-specific logic (finance, healthcare, etc.)
- SystemsReasoning: Complex system analysis — feedback loops, interdependencies, leverage points
The layer that monitors the platform's own cognitive processes.
- SecurityMonitoring: Real-time security event aggregation, threat correlation, automated incident response
- ContinuousLearning: Feedback loop — collects operational outcomes, generates learning insights, adapts system behaviour
- PerformanceMonitoring: KPI tracking (latency, throughput, resource utilisation) across all cognitive tasks
- SelfEvaluation:
MetacognitiveOversightComponent— quality, accuracy, and ethical alignment auditing of reasoning outputs - ReasoningTransparency: Generates explanations and justifications for AI decisions (explainability layer)
- CulturalAdaptation: Applies Hofstede's cultural dimensions to tailor agent interactions for global deployments
- Protocols: Manages ACP (AI Communication Protocol) and MCP (Metacognitive Protocol) for reliable inter-component communication
Where cognitive plans become real actions.
- MultiAgentOrchestration: Core engine for coordinating teams of specialised agents — task decomposition, agent selection, workflow execution, result synthesis
- ToolIntegration: Extensible tool framework (web search, data analysis, code execution, classification) — new tools added as adapters
- SecurityAgents: Automated incident response agents — immediate containment and remediation triggered by MetacognitiveLayer
- ActionPlanning: Multi-step plan generation for complex goal achievement
- DecisionExecution: Step-by-step plan execution with tool calls and state management
- HumanCollaboration: Human-in-the-loop infrastructure — agents pause for human review, approval, or intervention
- ConvenerAgents: Coordination agents that facilitate structured multi-party deliberation
The "front door" — all external clients (web, mobile, enterprise services) enter here.
| Application | What it exposes |
|---|---|
| Security | Zero-trust auth, risk scoring, compliance reporting |
| Customer Intelligence | Inquiry handling, conversation management, troubleshooting |
| Decision Support | Situation analysis, options generation, causal modelling |
| Knowledge Management | Document ingestion, knowledge base querying |
| Process Automation | Complex business process automation via agent workflows |
| Research & Analysis | Automated research, document synthesis, content generation |
| NIST Compliance | NIST RMF evidence collection, maturity assessment |
| Value Generation | Business value tracking and reporting |
- AIGovernance: Policy-as-code for AI operations, compliance dashboards
- CommunityPulse: Aggregates team/community signals for adaptive management
- LearningCatalyst: Active learning triggers based on system performance patterns
- UncertaintyQuantification: Surfaces confidence and uncertainty in AI outputs
A full-stack dashboard system (Next.js 15 frontend + .NET BFF).
- Widget marketplace: All UI functionality is a self-contained widget submitted through a governed review process (security scan → code signing → admin approval → registration)
- Personalised dashboards: Users compose their own layout from approved widgets;
DashboardLayoutServicepersists per-user configurations - PluginOrchestrator: "Sandwich pattern" gateway — every widget call is wrapped with auth, validation, and audit logging before reaching inner layers
- AgencyWidgets: Pre-built widgets for core mesh capabilities (Adaptive Balance, NIST RMF, agent status, etc.)
- Next.js frontend (
src/UILayer/web): Next.js 15 + React 19, Tailwind CSS, shadcn/ui, Storybook 8, D3 visualisations, i18n (en-US, fr-FR, de-DE), WCAG 2.1 AA accessibility, offline service worker
External clients (browsers, mobile, enterprise services)
|
v
UI Layer (Next.js 15 + .NET BFF)
|
v
Business Applications Layer (REST API — the platform's public surface)
/ | \
AgencyLayer MetacognitiveLayer ReasoningLayer
\ | /
FoundationLayer
(security, data, RAG, audit, secrets)
|
v
sluice gateway (all LLM calls route through here)
|
v
Azure OpenAI / model backends
All layers follow Hexagonal (Ports and Adapters) architecture: inner layers define typed ports (interfaces), outer layers provide adapters (implementations). Every infrastructure dependency is swappable without touching business logic.
cognitive-mesh/
├── src/
│ ├── FoundationLayer/ # Infra: security, data stores, RAG, audit
│ ├── ReasoningLayer/ # Cognitive engines: analytical, ethical, security, creative
│ ├── MetacognitiveLayer/ # Self-monitoring, learning, incident response
│ ├── AgencyLayer/ # Multi-agent orchestration, tools, automation
│ ├── BusinessApplications/ # REST API surface — the platform's front door
│ ├── UILayer/ # Plugin dashboard (Next.js + .NET BFF)
│ │ └── web/ # Next.js 15 frontend application
│ ├── MeshSimRuntime/ # Simulation runtime for testing mesh behaviour
│ └── ApiHost/ # ASP.NET Core hosting entry point
├── api/ # OpenAPI specs (convener-api.yaml)
├── tests/ # Unit + integration tests
├── cypress/ # E2E tests (cypress.config.ts)
├── k8s/ # Kubernetes manifests
├── infra/ # Infrastructure as code
├── docs/ # Architecture docs, runbooks, API versioning
├── examples/ # Usage examples per layer
├── .claude/ # Claude agent configuration (commands, hooks, rules)
├── AGENT_TEAMS.md # Agent team structure and responsibilities
├── CognitiveMesh.sln # .NET solution
├── Directory.Build.props # Shared build properties (version, targets)
└── docker-compose.yml # Local development stack
- .NET 10 SDK
- Node.js 20+ with npm (for UILayer/web)
- Docker (for local containerised run)
- Azure CLI (for cloud deployment)
- Access to sluice gateway endpoint (for LLM calls)
# Build the .NET solution
dotnet build CognitiveMesh.sln
# Run tests
dotnet test
# Run the API host locally
dotnet run --project src/ApiHost
# Run the UI frontend
cd src/UILayer/web
npm install
npm run dev # http://localhost:3000
npm run storybook # http://localhost:6006 (component docs)| Repo | Role |
|---|---|
| sluice | AI data plane — all LLM calls from cognitive-mesh route through sluice for routing, observability, and cost attribution |
| docket | Cost observability — attributes cognitive-mesh model spend via sluice OTEL spans; cost per agent run surfaced in Grafana |
| phoenix-flow | Project tracker — proxies task-routing decisions to cognitive-mesh for AI-assisted triage and planning |
| deck | Desktop ops tool — invokes cognitive-mesh agent teams for complex multi-step operations; surfaces agent status |
| retort | Agent scaffold — retort-based projects register their agent configs into the mesh; retort generates the .agentkit/ overlay |
| org-meta | Org intelligence — cognitive-mesh reads org-meta MCP context at session start; org-meta documents cognitive-mesh's cross-repo contracts |
| mystira-workspace | Mystira StoryGenerator delegates complex narrative planning to cognitive-mesh (AnalyticalReasoning for story coherence, CreativeReasoning for plot generation) |
- microsoft/semantic-kernel — .NET AI orchestration patterns
- microsoft/autogen — multi-agent conversation patterns
- mcowger/plexus — unified AI provider gateway (study reference for sluice integration patterns; note: no licence file — study only, do not copy code)
cognitive-mesh — a mesh network has no single point of failure; every node connects to every other. A cognitive mesh applies the same topology to AI: no single agent is the bottleneck, reasoning is distributed across specialised nodes, and the network routes around failures. The name deliberately avoids vendor lock-in — it describes the architecture, not the model or provider underneath it.
This project has not yet had a public release. Current development version: 0.0.1.