How does output propagation work with conditional edges

[kirankn8]

In the Strands Agents Graph documentation, the Feedback Loop example shows this pattern:

builder = GraphBuilder()
builder.add_node(draft_writer, "draft_writer")
builder.add_node(reviewer, "reviewer")
builder.add_node(publisher, "publisher")

builder.add_edge("draft_writer", "reviewer")
builder.add_edge("reviewer", "draft_writer", condition=needs_revision)
builder.add_edge("reviewer", "publisher", condition=is_approved)

The intended flow is:

draft_writer → reviewer → publisher (if approved)
                        → draft_writer (if needs revision, loop back)

My question is about what input publisher actually receives.
Since the only edge to publisher is reviewer → publisher, the publisher node would receive:

Original Task: ...

Inputs from previous nodes:

From reviewer:
  - Agent: [reviewer's assessment]

The draft_writer output is never directly passed to publisher because there is no draft_writer → publisher edge. The publisher only gets the reviewer's assessment, not the actual draft content.

Did I understand it correctly? If the publisher needs the actual draft content, would the correct approach be to add an additional conditional edge from draft_writer to publisher?

[mkmeral]

Yes, your understanding is correct. We check the results of all the incoming edges where condition is satisfied. See related code in graph.py

if you need the raw draft content directly available to the publisher, adding a conditional edge from draft_writer → publisher (with the same is_approved condition) is the right approach.

[kirankn8]

Thanks for replying @mkmeral . Adding a conditional edge from draft_writer → publisher solves the case where I want both outputs at the downstream node. But what about the case where I don't want the decision node's output at all?

In the below flow:

Here, the reviewer evaluates the draft and produces a structured review (e.g., {"approved": true, "feedback": "..."}). The quality check condition reads the reviewer's structured output to decide the route — if approved, flow goes to publisher; otherwise, back to draft_writer for revision.
The problem is that publisher only needs the draft content from draft_writer to publish it. It doesn't need the reviewer's assessment ("grammar is good, tone is appropriate, approved"). That's just noise in the publisher's input.
If I add a conditional edge draft_writer → publisher, then publisher receives both the draft and the review. But I want only the draft in publisher when quality check passes

[mkmeral]

Do you want that prompt to be passed around in LLM context? I think a better way could be passing the draft "as is" through shared state. That would allow you to save the draft to the state in draft writer. if/when a request comes to publisher, it can just publish the content in the shared state https://strandsagents.com/latest/documentation/docs/user-guide/concepts/multi-agent/multi-agent-patterns/#shared-state-across-multi-agent-patterns