strands-compose: Declarative YAML multi-agent orchestration for strands-agents

[galuszkm]

Hey everyone!

First off, I just want to say how much I absolutely love everything about strands. It's an absolute joy to work with!

As I've been building more complex systems, I noticed a bit of a gap in the ecosystem: we don't really have a standard, configurable way to build agent networks declaratively. We all know the power and scalability of IaC, Policy-as-Code, etc., so I won't preach to the choir about why configuration-driven setups are awesome. I did dig up some existing activity around this (like PR #935), which shows there's definitely interest and a ton of unrealized potential here.

To scratch my own itch and explore that potential, I started working on a project that fits exactly into this space: strands-compose.

What is strands-compose?

Simply put, it allows you to wire your entire agent system—models, prompts, tools, MCP servers, and orchestration topology (Delegate, Swarm, Graph) using single or multiple declarative YAML files.

Instead of writing repetitive plumbing code, you define your setup in YAML, and strands-compose loads it directly into a ready-to-use agent environment.

The Rules of Thumb

When building this, I stuck to two core principles to ensure it plays nicely with your ecosystem:

• Strands-first — If strands provides it, use it directly. No reinventing the wheel.
• Thin wrapper — Translate YAML → Python objects, then get out of the way.
  What you get back are plain, native strands objects

Just to be clear, I'm not here to advertise or claim any credit. This is a completely open-source project, and absolutely everyone is invited to jump in and collaborate!

I'd love your feedback! 🙏

I want to confront this idea with the reality:

• Does a declarative configuration approach appeal to you?
• What use cases do you think this would be great (or terrible) for?
• Are there any features you'd like to see?

I kindly invite you to check out the project. It's very well documented and I did my best to follow strands patterns as close as possible.