Streaming & Events

Agently normalizes model responses into an event stream. You can consume it synchronously or asynchronously:

• get_generator(...) for sync
• get_async_generator(...) for async

Event types

Common events (OpenAICompatible as reference):

• delta: text delta
• reasoning_delta: reasoning delta
• tool_calls: tool call chunks
• done: final text
• original_delta / original_done: raw provider stream
• meta: usage / finish_reason, etc.
• extra: extra fields
• error: request or parsing errors

Text stream only (delta)

for chunk in response.get_generator(type="delta"):
  print(chunk, end="", flush=True)

Filter specific events

for event, value in response.get_generator(
  type="specific",
  specific=["reasoning_delta", "tool_calls", "done"],
):
  print(event, value)

Full event stream

for event, value in response.get_generator(type="all"):
  print(event, value)

Structured streaming (instant / streaming_parse)

When output uses Agently Output Format, you can stream structured updates:

for item in response.get_generator(type="instant"):
  print(item.path, item.delta, item.is_complete)

instant is equivalent to streaming_parse. Path style is controlled by response.streaming_parse_path_style (dot or slash).

See also: Instant structured streaming.