The enterprise sector is showing increasing interest in the concept of Agentic AI—systems where neural networks act not just as chat conversationalists, but as autonomous digital workers. However, strict security requirements and regulatory restrictions make it impossible to use public cloud services like solutions from OpenAI or Anthropic when working with critical documents. The only way out is full on-premise deployment within an isolated perimeter.
Infrastructure Layer: Kubernetes and Complete Isolation
To simplify administration for enterprise DevOps teams, the Archangel platform was packaged into standard Helm charts, avoiding non-standard installers. This ensures predictable deployments and easy horizontal scaling.
Key principles of the closed infrastructure setup:
- Air-gapped environment: All container images and local models (Llama, Mistral families, etc.) are stored exclusively in internal secure registries (Harbor or Nexus).
- State management: StatefulSet manifests allow deploying databases and message brokers inside the cluster or integrating the platform with existing corporate PostgreSQL and Redis storages.
- Network policies: At the Kubernetes network policy level (using CNI plugins with policy support, such as Calico or Cilium), any outgoing connections to external networks are blocked at the root.
Agent Orchestration via Event-Driven Architecture
To prevent hallucinations, tasks are not assigned to a single universal model, but are distributed among highly specialized agents. Their interaction is organized via an event-driven architecture (EDA) over an internal data bus.
Example of draft contract processing:
- Ingestion Agent: Extracts and cleanses document text, then sends a readiness notification to the bus.
- Legal Agent: Performs semantic auditing, identifies potential legal conflicts or penalty risks, and emits the corresponding event forward.
- Finance Agent: Receives the signal from the legal agent and calculates financial risks based on the organization's internal statistics.
- Integration Agent: Aggregates the gathered information and generates a structured response for internal systems (ERP, CRM, or 1C).
Thanks to the asynchronous scheme, agents do not block each other, and when incoming document volume increases, it is sufficient to scale the number of pods for the most loaded components.
Local RAG and Vector Operations
To enable agents to rely on company regulations and archives, a Retrieval-Augmented Generation (RAG) architecture implemented entirely on-premises is used:
- Internal embeddings: Text vectorization is performed by lightweight open-source models on cluster's own GPU nodes without sending data outside.
- Isolated vector databases: Vector representations of regulatory documents are stored in Qdrant, Milvus, or pgvector inside the perimeter.
- Local updates: Knowledge base updates happen entirely within the company—when new documents are added, the system re-indexes data locally without any cloud synchronization.
Implementing Agentic AI for the enterprise segment is primarily a challenge of network isolation and well-thought-out service interaction architecture. Leveraging an event-driven approach and local models enables an autonomous system operating without a single external network request.