Agently Docs

EN

中文

EN

中文

Appearance

TriggerFlow Overview ​

TriggerFlow is not “a more complicated request call”. It is Agently's workflow layer for explicit stages, concurrency, waiting, resume, and runtime events.

When to read this ​

  • Your problem no longer fits one request
  • You need stages, branches, concurrency, or wait/resume
  • You want the workflow itself to be observable, testable, and reviewable

What you will learn ​

  • When to upgrade into TriggerFlow
  • The signal-driven mental model behind TriggerFlow
  • How blueprint, execution, and runtime state relate to each other

TIP

Real TriggerFlow practice should default to Async First: prefer async chunks, async_start(...), get_async_runtime_stream(...), and response.get_async_generator(type="instant") inside chunks when model work needs structured streaming.

When you should upgrade ​

If your problem already has these traits, stop forcing it into one request:

  • the same task runs through several explicit stages
  • you need conditional routing or branching
  • you need concurrency across several items
  • you need to pause and continue after external input
  • you need runtime stream or execution save/load

TriggerFlow mental model ​

Minimal example ​

python
from agently import TriggerFlow, TriggerFlowRuntimeData

flow = TriggerFlow()

@flow.chunk("normalize")
async def normalize(data: TriggerFlowRuntimeData):
    return str(data.value).strip()

@flow.chunk("greet")
async def greet(data: TriggerFlowRuntimeData):
    return f"Hello, {data.value}"

flow.to(normalize).to(greet).end()
print(flow.start(" Agently "))

The example above keeps a sync entry on purpose so the structure stays easy to read. As soon as this becomes a real service or observable workflow, the recommended upgrade is:

  • async chunk handlers
  • execution.async_start(...)
  • execution.get_async_runtime_stream(...)
  • response.get_async_generator(type="instant") for model steps
  1. Concepts
  2. Basic Flow
  3. Events and Signals
  4. Data and Resources
  5. Runtime Stream
  6. From Token Output to Live Signals
  7. Blueprint

Common mistakes ​

  • Assuming you need TriggerFlow when the real problem is only unstable output control
  • Treating TriggerFlow like a static DAG instead of a signal-driven runtime
  • Designing nodes and shared state before clarifying the workflow owner layer

Next ​

  • agently-triggerflow
  • agently-playbook