Clarify Partial Message Streaming and Output Placement

[matoushavlena]

There is currently no clear support in A2A for streaming partial Messages (for example, token-by-token). Unlike Artifacts, which support append and lastChunk, Messages are treated as atomic. There is no protocol-defined way to send or reassemble a streamed message.

This creates challenges for client developers:

• Token-level streaming is not reliably supported across implementations
• Workarounds like sending multiple partial Messages pollute task history
• It is unclear how to reconstruct the actual output the agent intended to send

A2A would benefit from a standardized approach to message streaming so that clients can provide real-time feedback without fragile hacks.

Related problem: ambiguous output location

It is also not clearly defined where the agent’s primary output is supposed to appear. The response could be:

• a Message
• a TaskStatusUpdateEvent.status.message
• an Artifact

If more than one of these is used in the same task, it becomes unclear which is authoritative:

• Is the intended output the last message?
• The message in the final task status?
• The last artifact?
• What if the agent sends multiple artifacts (like images) along with a message (like a caption)?

This ambiguity makes it difficult for agents and client applications to know what to act on or display.

Suggestions

• Define semantics for partial Message streaming, or formally recommend using Artifacts for token-level output
• Clarify how clients and agents should determine which output is the primary one (message, artifact, or task status)

Would be great to hear how others are approaching this or whether any conventions are emerging.

Related issues/discussions

• A better way to handle Agent stream output for client display #774