← All Articles

A2A Bidirectional Streaming Proposal: gRPC Gets Native Agent Communication

The Proposal

HungYangChang opens PR #1549 proposing native bidirectional streaming support over gRPC for the Agent-to-Agent Protocol. This is a breaking change (marked feat!) that would fundamentally alter how agents communicate in real-time scenarios.

Currently, A2A uses a request-response pattern where agents send messages and poll for responses. Bidirectional streaming would enable agents to maintain persistent connections with simultaneous send/receive capabilities — essential for collaborative, multi-turn interactions between autonomous agents.

Author Background

HungYangChang is contributing to the A2A ecosystem with a focus on performance-critical use cases. This proposal emerges from practical experience implementing agent communication where polling latency becomes a bottleneck.

Why gRPC Bidirectional Streaming?

The A2A protocol currently supports multiple transport mechanisms, but lacks native streaming for real-time agent interactions. gRPC bidirectional streaming provides:

Capability	Current A2A	With Bidirectional Streaming
Message Latency	Poll-based (100ms+ typical)	Sub-millisecond push
Connection Overhead	New connection per message	Single persistent stream
Backpressure	Client-managed	Native gRPC flow control
Interruption	Requires separate cancel request	Stream close signals interruption

Technical Details

The proposal would add a new streaming RPC to the A2A protocol definition:

service AgentService {
  // Existing unary methods
  rpc SendMessage(SendMessageRequest) returns (SendMessageResponse);
  
  // New bidirectional streaming
  rpc StreamMessages(stream AgentMessage) returns (stream AgentMessage);
}

Key design considerations in the proposal:

• Message Framing: Each message in the stream includes task context, enabling multiplexed conversations over a single connection
• Error Propagation: gRPC status codes map to A2A error types, with stream-level errors distinct from message-level errors
• Graceful Degradation: Implementations can fall back to unary RPCs if streaming isn't supported by the counterparty

Use Cases Enabled

1. Collaborative Reasoning

Two agents working on a complex problem can exchange intermediate results in real-time, building on each other's partial solutions without the latency of polling.

2. Human-in-the-Loop Supervision

A supervisory agent monitoring multiple worker agents can receive status updates instantly and inject corrections without waiting for poll intervals.

3. Agentic Workflows with Dependencies

When Agent B's work depends on Agent A's output, streaming enables A to push results as they become available, rather than B polling repeatedly.

Challenges to Address

The PR discussion highlights several open questions:

• Connection Lifecycle: How long should streams stay open? What triggers reconnection?
• Load Balancing: Long-lived gRPC streams complicate load balancer behavior; the proposal needs to address connection distribution
• State Management: If a stream disconnects mid-conversation, how is state reconciled?
• Security: Persistent connections increase attack surface; the proposal should address authentication refresh and connection auditing

Ecosystem Implications

If adopted, bidirectional streaming would represent a significant evolution in A2A's capabilities:

• SDK Updates: All official A2A SDKs would need streaming support
• Infrastructure: Deployment patterns shift — agents need to handle long-lived connections
• Observability: Monitoring tools need stream-aware tracing, not just request/response logging

This proposal is still in active discussion. The A2A maintainers are evaluating whether streaming should be a core protocol feature or remain an optional extension.