ADK Workflow Rules That Bite

The failure modes that account for most broken ADK workflows. Each one is a silent or confusing failure, not a clear error — that is why they are collected here rather than left to be discovered.

Type everything with Pydantic models

Use a BaseModel for node inputs, node outputs, LLM output_schema, RequestInput.response_schema, and structured state values. dict[str, Any] gives up validation, IDE completion, and the automatic dict-to-model conversion that FunctionNode performs from type hints.

# Loses validation and downstream typing
def lookup_flights(node_input: dict[str, Any]) -> dict[str, Any]:
  return {'flight_cost': 500, 'details': 'Economy'}


# Validated on the way in and on the way out
class FlightInfo(BaseModel):
  flight_cost: int
  details: str


def lookup_flights(node_input: Itinerary) -> FlightInfo:
  return FlightInfo(flight_cost=500, details='Economy')

event.output is plumbing; event.content is the UI

The web UI renders event.content only. A node that produces something a human should read must emit it as a message, in addition to the output the next node consumes.

def final_output(node_input: str):
  yield Event(message=node_input)  # rendered in the web UI
  yield Event(output=node_input)   # passed to the downstream node

Event(message=...) is a constructor convenience that writes event.content. Stream partial text with Event(message='chunk', partial=True). LLM agents emit their content events automatically; function nodes do not.

Write state through Event(state=...), not ctx.state[key] = ...

State passed to the Event constructor lands in event.actions.state_delta, so it is part of event history and survives the replay that non-resumable human-in-the-loop performs. A direct ctx.state mutation is a side effect that replay does not reproduce.

# Recorded in event history
def save(node_input: str):
  return Event(output=node_input, state={'user_request': node_input})


# Lost on replay
def save(ctx: Context, node_input: str) -> str:
  ctx.state['user_request'] = node_input
  return node_input

Reading is always ctx.state[...].

At most one output event per node execution

A node may yield many events, but only one of them may carry output. Two output events are merged into a list, which silently changes the downstream node's node_input type from str to list[str]. The same limit applies to route, except there a second routed event raises ValueError.

def my_node(node_input: str):
  yield Event(message='Processing...')   # display only
  yield Event(state={'status': 'done'})  # state only
  yield Event(output='final result')     # the one output

This holds for function nodes, LLM agent nodes, and nested workflows alike.

A function either yields or returns — never both

Python turns any function containing yield into a generator and discards its return value. Mixing the two loses the output with no error.

# Generator: every event is yielded
def my_node(node_input: str):
  yield Event(state={'key': 'value'})
  yield Event(output='result')


# Plain function: one return, which may be an Event or a bare value
def my_node(node_input: str):
  return Event(output='result', state={'key': 'value'})


# Broken: the return is silently ignored
def my_node(node_input: str):
  yield Event(state={'key': 'value'})
  return Event(output='result')

{var} in an instruction reads state, never node_input

Instruction placeholders resolve against ctx.state only. node_input is delivered to the model as the user message, so {node_input} is just a missing state key and raises KeyError at call time.

# KeyError: 'node_input' is not a state key
Agent(name='summarizer', instruction='Summarize this: {node_input}')

# The predecessor's output is already the user message
Agent(name='summarizer', instruction='Summarize the text in one sentence.')

# Anything else the instruction needs must be in state first
Agent(name='writer', instruction='Write about "{topic}". Feedback: {feedback?}')

{var?} substitutes an empty string when the key is absent; {var} raises.

A chat-mode agent can only follow START

Graph validation rejects an edge into an LlmAgent with mode='chat' from any node other than START, because a chat agent reads conversational history rather than a node input. Set mode='single_turn' (the default for an auto-wrapped agent) or mode='task' on agents used mid-graph.

Everything in event.output must be JSON-serializable

• FunctionNode converts a returned BaseModel with model_dump(), so returning a model is safe. A types.Content or any other non-serializable object is not.
• An LLM agent with output_schema validates then dumps, so ctx.state[output_key] is a plain dict, never a model instance. Read it with data['field'], or rebuild the model with MyModel(**data).
• This bites hardest at a JoinNode, which parks partial inputs in session state while waiting. If a predecessor is an LLM agent without output_schema, the parked value is a types.Content and DatabaseSessionService raises TypeError on write.

Give every loop iteration its own interrupt_id

A node that yields RequestInput inside a loop must vary the interrupt_id per iteration. Reusing one id makes event-based state reconstruction match an earlier iteration's answer to the current interrupt, which restarts the loop forever.

review_count = ctx.state.get('review_count', 0)
yield RequestInput(
    interrupt_id=f'review_{review_count}',
    message='Approve?',
)

Increment the counter through Event(state={'review_count': review_count + 1}) when the response arrives.