Agent Architecture: IO + Inference Separation

Tagentacle adopts a clean separation between Agent Nodes (context engineering + agentic loop) and Inference Nodes (stateless LLM gateway).

Agent Node = Complete Agentic Loop

An Agent Node is a single Pkg that owns the entire agentic loop internally:

• Subscribe to Topics → receive user messages / events
• Manage the context window (message queue, context engineering)
• Call Inference Node's Service for LLM completion
• Parse tool_calls → execute tools via MCP Session (Streamable HTTP direct connection) → backfill results → re-infer

This loop is a tightly-coupled sequential control flow (like ROS 2's nav2 stack) and should not be split across multiple Nodes.

Inference Node = Stateless LLM Gateway

A separate Pkg (official example at org level, not part of the core library) that provides:

• A Service (e.g., /inference/chat) accepting OpenAI-compatible format: {model, messages, tools?, temperature?}
• Returns standard completion: {choices: [{message: {role, content, tool_calls?}}]}
• Multiple Agent Nodes can call the same Inference Node concurrently

Data Flow

UI Node ──publish──▶ /chat/input ──▶ Agent Node (agentic loop)
                                        │
                                        ├─ call_service("/inference/chat") ──▶ Inference Node ──▶ OpenRouter/OpenAI
                                        │                                           │
                                        │◀── completion (with tool_calls) ◀─────────┘
                                        │
                                        ├─ MCP Session (HTTP) ──▶ Tool Server Node
                                        │◀── tool result ◀────────┘
                                        │
                                        └─ publish ──▶ /chat/output ──▶ UI Node