Architecture Intermediate ⏱️ 10 min

Architecture Patterns Overview

Patterns are reusable answers to recurring design questions. This page is the map; each pattern links to a full lesson with working code.

Single agent with tools

The simplest useful shape: one agent, a focused instruction set, and a handful of function tools. Reach for this when a task fits in one reasoning loop.

Single agent with function tools

flowchart LR
    U[User] --> A[Agent]
    A -->|calls| T1[get_inventory]
    A -->|calls| T2[create_order]
    T1 --> A
    T2 --> A
    A --> U

One agent, deterministic tools, one conversation thread.

Supervisor / orchestrator

When work spans multiple specialities, a supervisor routes to specialist agents and aggregates their results. This keeps each agent's context small and its instructions sharp.

Supervisor routing to specialists

flowchart TD
    U[User] --> S[Supervisor Agent]
    S --> A[Billing Agent]
    S --> B[Logistics Agent]
    S --> C[Support Agent]
    A --> S
    B --> S
    C --> S
    S --> U
    classDef hi fill:#6366f1,stroke:#4f46e5,color:#fff;
    class S hi;

A coordinator delegates to focused sub-agents and merges the outcome.

⚠️

Don't over-orchestrate

Multi-agent systems add latency and cost. Start with one agent; introduce a supervisor only when a single agent's instructions become unmanageable or its tools conflict.

Remote tools via MCP

Expose capabilities as MCP servers so any agent can discover and call them over a network boundary — hosted on Function Apps, Logic Apps, or a FastMCP service, and secured behind a gateway with OAuth.

Agents calling secured remote MCP servers

flowchart LR
    A1[Agent A] --> G["Agent Gateway
OAuth · rate limit"]
    A2[Agent B] --> G
    G --> R[(Agent Registry)]
    G --> M1[MCP · Function App]
    G --> M2[MCP · Logic App]
    G --> M3[FastMCP Service]
    classDef hi fill:#6366f1,stroke:#4f46e5,color:#fff;
    class G hi;

A gateway centralises auth, rate limiting, and discovery for remote tools.

Each pattern above has a dedicated lesson with runnable code, deployment steps, and the trade-offs that matter in production.