Agently Docs

EN

中文

EN

中文

Appearance

Instant Structured Streaming ​

Instant solves this problem: you want early feedback, but you do not want to fall back to a noisy stream of raw text tokens.

When to read this ​

  • You already define output with output()
  • You want realtime UI, progressive summaries, or early downstream triggers
  • You do not want to wait for the full JSON-like result before starting to work

What you will learn ​

  • Why structured streaming requires output() first
  • What path, wildcard_path, and delta mean
  • When to fix one response and consume both stream and final data from it
  • Why get_async_generator(type="instant") should be the default in services and TriggerFlow

Instant data flow ​

Minimal example ​

python
from agently import Agently

agent = Agently.create_agent()

response = (
    agent
    .input("Explain recursion, give one definition and two tips")
    .output(
        {
            "definition": (str, "One-line definition"),
            "tips": [(str, "Short tip")],
        }
    )
    .get_response()
)

for item in response.result.get_generator(type="instant"):
    if item.path == "definition" and item.delta:
        print(item.delta, end="", flush=True)
    if item.wildcard_path == "tips[*]" and item.delta:
        print(item.delta, end="", flush=True)

print()
print(response.result.get_data())

Async First version ​

If this runs inside a service, async worker, SSE endpoint, or TriggerFlow chunk, prefer:

python
async for item in response.get_async_generator(type="instant"):
    if item.is_complete and item.wildcard_path == "tips[*]":
        print(item.path, item.value)

final_data = await response.async_get_data()

This is especially useful when:

  • one response updates UI and also feeds later logic
  • completed fields should be forwarded into TriggerFlow
  • several async consumers reuse the same response

Reading path and wildcard_path ​

  • path: the concrete field path, such as tips[0]
  • wildcard_path: the pattern form, such as tips[*]
  • delta: the incremental text chunk for that field

For list-like UIs, wildcard_path is often the better matching surface.

Why get_response() matters here ​

If you want both streaming updates and final structured data, fix one response first:

  • streaming stage: response.get_async_generator(type="instant")
  • final stage: await response.async_get_data()

That keeps one model request and several result views.

Common mistakes ​

  • Trying structured streaming before defining output()
  • Reconstructing fields manually from delta text
  • Re-requesting the same result just to get the final data after streaming

Next ​

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