I have been building multi agent systems and the hardest part in practice has been shared state once agents are split across processes or services. Retries overwrite things, handoffs get messy, and crashes make recovery painful.

State inside a single workflow or graph works well. Outside of that boundary, I have not found many patterns that hold up in production.

For context, I built a small open source, self hosted experiment to explore this idea. Repo: https://github.com/MehulG/memX

I am curious how others are handling this today. Are people centralizing state, rebuilding from logs, or avoiding shared state altogether? What has actually worked for you?

If anyone is actively running multi agent workflows and is open to trying this on a single flow, I am happy to help with setup and learn from the experience.