This is a great example of why multi-agent systems need to be treated like distributed systems, not just bigger prompts. Once agents share state, that state becomes a trust boundary. If provenance, validation, and ownership aren't built into the shared memory, one compromised agent can quietly influence every downstream decision. It feels very similar to event-driven architectures every message needs context about where it came from before another service acts on it. I think we'll see "state hygiene" become as fundamental to agentic AI as input validation is for traditional applications.