Hello everyone,

First, I want to say how grateful I am for this community. I’ve learned so much just reading through these threads. I’m relatively new to deploying agent frameworks in a serious way, and I’m trying to be very mindful of security from the start in my homelab setup.

I’m currently evaluating two open-source multi-agent frameworks: Goose (from Block) and Microsoft’s Autogen. My primary use case involves a Python-based workflow for automating some data analysis and report generation, all running inside Docker containers on a private network. While functionality is important, my deciding factor is likely going to be which one has a fundamentally better security posture I can build upon.

I’ve been reading the docs for both and have some initial thoughts, but I’d really appreciate the expertise here to help me see what I might be missing. My concerns naturally lean towards container security, credential isolation, and the risks introduced by the extension/agent model.

Here’s my layman’s breakdown of where I see the security considerations for each:

* **Goose (Block):**
* Its extension model is very powerful, but it makes me nervous. Each extension runs in its own sandbox, which is good, but the documentation mentions the “local execution context” for trusted extensions. I worry about properly configuring that boundary.
* I like that it’s open-source and from a company (Block) with a security focus, which should help the audit story. But being a newer project, has the codebase been scrutinized enough?
* How does credential handling work for extensions that need API keys? Is there a secure injection mechanism, or do they risk being exposed in the agent’s memory space?

* **Microsoft Autogen:**
* It feels more “academic” in its approach, with a heavy focus on the agent conversation patterns. The security seems a bit more… implicit?
* It’s also open-source, but under the Microsoft umbrella. Does that lead to more eyes on the code, or does it get lost in a huge org?
* I’ve seen that you can deploy agents as Docker containers, which is my plan. But I’m less clear on how it manages inter-agent communication security and if there’s any built-in principle of least privilege for what tools an agent can access.

My gut feeling is that Goose, with its explicit security model, might be designed more defensively. But Autogen’s maturity and backing could mean more robust, battle-tested practices. I’m probably over-explaining my own setup here, but I think context matters: I’ll have these agents handling slightly sensitive personal data, so even in a homelab, I want to get the foundations right.

Has anyone here conducted a deeper security analysis of either framework, specifically looking at their isolation models, supply chain security (dependencies!), or how they handle secrets? Any practical experiences or red flags would be incredibly helpful.

- Liam