Skip to main content

Agentic Graph Operations: Agents as First-Class Citizens

In Reaktor, AI agents are not just external tools; they are full participants in the graph runtime. This "Agent-as-Node" architecture allows for a more powerful, granular, and secure agent runtime than standard Model Context Protocol (MCP) integrations.

Agent-as-Node Architecture​

An agent in Reaktor is a specialized graph node (AgentOperatorNode) that possesses several key characteristics:

  • An Identity: Every agent is bound to a Reaktor Auth principal (service account) with specific RBAC roles and permissions.
  • Typed Port Access: The agent's ConsumerPorts and ProviderPorts define exactly what data it can read and write, which is enforced at the graph level.
  • An Operation Log: Every port invocation, state mutation, and database query is traced and visible in the Blueprint editor in real-time.
  • Rollback Capability: Agent operations can be wrapped in a transaction scope, allowing for automatic rollback if a multi-step plan fails.
  • Scoped Context Window: The agent's LLM context is populated directly from the graph—the topology of reachable nodes, the schemas of accessible ports, and the recent telemetry of connected edges.

Auth-Governed Agent Operations​

Every agent operation is governed by Reaktor Auth's granular RBAC system. An agent can only access ports and nodes for which it is authorized.

  • Database Queries: Access RepositoryNode's ProviderPorts with db:read and db:write permissions.
  • Graph Traversal: Uses the GraphInspector interface for read-only traversal with graph:read permissions.
  • Graph Mutation: Add nodes, wire ports, and hot-swap implementations with graph:write and graph:deploy permissions.
  • Service Invocation: Call external APIs through a ServiceNode with service:invoke permissions.
  • Deployment: Push JS bundles to R2 and trigger k3s rolling updates with deploy:js_mobile or deploy:server permissions.

Why This Is Better Than MCP​

While Reaktor supports MCP for external interoperability, its graph-native agent model solves several fundamental limitations:

  • Relationship-Aware: Agents operate on a graph with typed edges, allowing them to understand relationships and dependency ordering.
  • Built-in Auth: Agents have RBAC identities, ensuring secure tool access.
  • Transaction Support: Multi-step sequences can fail with an automatic rollback.
  • Observability: Every operation flows through Reaktor telemetry and is visible in a unified trace.
  • Context-Rich: The agent's context is populated from the graph itself, which acts as the tool documentation.