Transports

Agent Runtimes supports multiple transport protocols for communicating with AI agents. Each transport has different characteristics suited for various use cases.

Protocol Comparison

Feature	AG-UI	Vercel AI	ACP	A2A
Transport	HTTP/SSE	HTTP/SSE	WebSocket	HTTP/SSE
Streaming	✅	✅	✅	✅
Tool Calling	✅	✅	✅	✅
Per-Request Model Selection	✅	✅	✅	❌
Bidirectional Communication	❌	❌	✅	❌
Human-in-the-Loop	✅	❌	✅	❌
Inter-Agent Communication	❌	❌	❌	✅

AG-UI (Agent UI Protocol)

AG-UI is a lightweight protocol optimized for building interactive AI chat interfaces.

Features:

• Simple HTTP/SSE-based communication
• Built-in support for streaming responses
• Tool calling with human-in-the-loop approval
• Native support for frontend components

Use Cases:

• Web chat applications
• Interactive AI assistants
• Applications requiring tool approval workflows

Endpoint: POST /api/v1/ag-ui/agents/{agent_id}/runs

// Frontend usage
const adapter = new AGUIAdapter({
  type: 'ag-ui',
  baseUrl: 'http://localhost:8888/api/v1/ag-ui/agents/demo',
});

Vercel AI Protocol

Compatible with Vercel's AI SDK, enabling seamless integration with Next.js and React applications.

Features:

• Vercel AI SDK compatibility
• Server-Sent Events (SSE) streaming
• Tool calling support
• Easy integration with existing Vercel projects

Use Cases:

• Next.js applications
• Projects already using Vercel AI SDK
• Quick prototyping with familiar APIs

Endpoint: POST /api/v1/vercel-ai/agents/{agent_id}/chat

// Frontend usage
const adapter = new VercelAIAdapter({
  type: 'vercel-ai',
  baseUrl: 'http://localhost:8888/api/v1/vercel-ai/agents/demo',
});

ACP (Agent Client Protocol)

ACP is a WebSocket-based protocol providing rich bidirectional communication.

Features:

• Full-duplex WebSocket communication
• Session management
• Permission request/grant workflow
• Real-time streaming updates
• Cancellation support

Use Cases:

• Long-running agent tasks
• Applications requiring bidirectional updates
• Complex workflows with permission handling
• Real-time collaborative experiences

Endpoint: WS /api/v1/acp/ws/{agent_id}

// Frontend usage
const adapter = new ACPAdapter({
  type: 'acp',
  baseUrl: 'ws://localhost:8888/api/v1/acp/ws/demo',
});

A2A (Agent-to-Agent Protocol)

A2A enables communication between AI agents using the Google A2A standard via fasta2a.

Features:

• Inter-agent communication
• Task delegation and result aggregation
• Agent discovery via agent cards
• Compatible with pydantic-ai agents

Use Cases:

• Multi-agent systems
• Orchestrating multiple specialized agents
• Agent collaboration workflows

Endpoint: POST /api/v1/a2a/agents/{agent_id}/

// Frontend usage
const adapter = new A2AAdapter({
  type: 'a2a',
  baseUrl: 'http://localhost:8888/api/v1/a2a/agents/demo/',
});

A2A Limitations

Per-request model selection is not supported with A2A.

The A2A protocol uses pydantic-ai's to_a2a() method which binds the model at agent creation time. The model cannot be changed per-request. If you need per-request model selection, use AG-UI, Vercel AI, or ACP protocols instead.

See the A2A Limitations section for more details.

A2A Limitations

Per-Request Model Selection

The A2A protocol does not support changing the model on a per-request basis. This is due to:

1. Architecture: A2A uses fasta2a which converts pydantic-ai agents via to_a2a(). The model is fixed at agent creation.

2. Design Philosophy: A2A treats agents as services with fixed capabilities. The model is part of the agent's identity.

Workarounds:

1. Multiple Agents: Deploy separate A2A agents for each model:

   # Create agent with GPT-4
   gpt4_agent = Agent("openai:gpt-4o")
   register_a2a_agent(PydanticAIAdapter(gpt4_agent), card_gpt4)
   
   # Create agent with Claude
   claude_agent = Agent("anthropic:claude-sonnet-4-5")
   register_a2a_agent(PydanticAIAdapter(claude_agent), card_claude)

2. Use Different Protocol: Switch to AG-UI, Vercel AI, or ACP for per-request model selection.

Choosing a Transport

Requirement	Recommended Transport
Simple web chat	AG-UI or Vercel AI
Next.js / Vercel project	Vercel AI
Real-time bidirectional updates	ACP
Human-in-the-loop workflows	ACP or AG-UI
Multi-agent orchestration	A2A
Per-request model switching	AG-UI, Vercel AI, or ACP