Agently Docs

EN

中文

EN

中文

Appearance

Streaming Responses and Event Types ​

Agently normalizes model output into one event stream so you do not have to bind your application to each provider's raw streaming shape.

When to read this ​

  • You already understand get_response()
  • You want to connect streaming events to UI, logs, or downstream logic
  • You want to distinguish delta, meta, tool_calls, and instant

What you will learn ​

  • Why Agently normalizes provider streams first
  • What the common event types mean
  • When to use text streaming and when to use structured streaming
  • Why async generators should be the default in production

Event flow ​

instant is not a separate system. It is a structured view built on top of the same normalized event layer.

Common event types ​

  • delta: text delta
  • reasoning_delta: reasoning delta
  • tool_calls: tool-call fragments
  • meta: usage, finish_reason, and similar metadata
  • done: final result
  • extra: extended fields
  • error: request or parsing error
  • instant / streaming_parse: structured streaming events

Async First recommendation ​

If you are inside a service runtime, SSE endpoint, WebSocket handler, TriggerFlow chunk, or any overlapping-request workload, prefer async generators:

python
async for item in response.get_async_generator(type="instant"):
    print(item.path, item.value)

Why:

  • avoids blocking sync bridges inside an event loop
  • fits async service integration better
  • makes it easier to forward structured stream events into TriggerFlow or runtime stream

Text-only streaming ​

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

Good fit:

  • typewriter-style UI
  • simple logs

Filter only the events you care about ​

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

Structured streaming ​

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

Good fit:

  • the output schema is already explicit
  • you want to update one field or list item while generation continues

In production, the more recommended form is:

python
async for item in response.get_async_generator(type="instant"):
    if item.is_complete:
        print(item.path, item.value)

Common mistakes ​

  • Using only delta when the real need is structured field updates
  • Exposing raw event noise to the frontend instead of turning it into business semantics first
  • Starting several consumers before you have fixed one reusable response

Next ​

  • agently-model-response
  • agently-output-control