agents-go

Streaming

agents.RunStreamed runs the same agent loop as Run but emits events while it executes — raw model deltas, completed run items, and agent switches.

sr := agents.RunStreamed(ctx, agent, "Tell me 5 jokes.", agents.RunOptions{ModelProvider: provider})

for event, err := range sr.Events() {
	if err != nil {
		log.Fatal(err) // terminal error; iteration stops
	}
	switch ev := event.(type) {
	case *agents.RawResponsesStreamEvent:
		// token-by-token deltas, verbatim from the Responses API
		if ev.Data.Type == "response.output_text.delta" {
			fmt.Print(ev.Data.Delta) // union exposes the variant fields flattened
		}
	case *agents.RunItemStreamEvent:
		fmt.Printf("\n[%s]\n", ev.Name) // e.g. message_output_created, tool_called, tool_output
	case *agents.AgentUpdatedStreamEvent:
		fmt.Printf("\n[now talking to %s]\n", ev.NewAgent.Name)
	}
}

res, err := sr.FinalResult() // the completed RunResult, same as a non-streamed run

Events() is an iter.Seq2[StreamEvent, error] — Go’s range-over-func replaces Python’s async for.

Event types

Raw response events

*RawResponsesStreamEvent passes through every OpenAI Responses streaming event (response.created, response.output_text.delta, …). Use these for token-level UI streaming.

Run item events and agent events

*RunItemStreamEvent fires when an item is complete (a full message, a tool call, a tool result) — the right granularity for “Fetching the weather…”-style progress, ignoring per-token noise. Event names mirror Python: message_output_created, tool_called, tool_output, handoff_requested, handoff_occured, reasoning_item_created.

*AgentUpdatedStreamEvent fires when a handoff switches the active agent.

Semantics

• Events are delivered on a buffered channel; the producing goroutine blocks when you stop consuming. If you break out of the loop early, cancel the run’s context — otherwise the run goroutine leaks waiting to deliver.
• A failing run yields one terminal error from the iterator, and FinalResult() returns the same error.
• Everything else works as in Run: sessions are saved on completion, guardrails fire (input guardrails run synchronously before the first model call in streamed mode), tracing records the same spans.
• A streamed run can pause for tool approval: iterate to the end, then check FinalResult()’s Interruptions/State and resume with agents.ResumeRun (the resumed run is non-streamed).

This site is open source. Improve this page.