09_graph

09 — Graph Pipeline

A fixed pipeline — nodes execute in the order you define, no surprises.

What this shows

• mode: graph in orchestrations: — a deterministic DAG with explicit edges
• edges: list — each edge is a from: -> to: pair that locks in execution order
• entry_name: — the starting node (must have no incoming edges)
• The difference between graph and swarm: you decide the order, not the agents

How it works

A graph pipeline executes nodes in topological order. The output of each node is passed as context to the next.

content_strategist ──▶ content_writer ──▶ copy_editor

In YAML, edges are explicit from: / to: pairs:

orchestrations:
  blog_pipeline:
    mode: graph
    entry_name: content_strategist
    edges:
      - from: content_strategist
        to: content_writer
      - from: content_writer
        to: copy_editor

That's it — three agents, two edges, one pipeline. strands-compose builds the DAG and runs each node in sequence.

Good to know

Graph vs Swarm vs Delegate. Use graph when the execution order is fixed and known upfront — e.g. content pipelines, multi-stage validation, ETL. Use swarm for collaborative back-and-forth. Use delegate for a coordinator that picks tools on the fly.

Nodes at the same depth can run in parallel. If two nodes share no dependency, strands executes them concurrently — but that's handled for you.

Conditional edges are supported too — see example 13 for that.

Prerequisites

• AWS credentials configured (aws configure or environment variables)
• Dependencies installed: uv sync

Run

uv run python examples/09_graph/main.py

Try these prompts

• Write a blog post about why Rust is gaining popularity among Python developers.
• Create a post about the importance of code review in software teams.
• Write about how large language models are changing software development.

Advanced topic — suppress default callback logging

Strands agents log actions to the console through their default callback_handler. If you want cleaner example output, set the handler to null in agent_kwargs for any agent:

agents:
  my_agent:
    agent_kwargs:
      callback_handler: null # or ~