Let me show you how isolation works under the hood. Each agent on your Gateway gets its own directory at ~/.openclaw/agents/<agentId>/. Inside that directory, the agent has its own session history, MEMORY.md, and configuration. No sharing by default.
What happens if one agent gets compromised? If you enable Docker sandboxing, each agent runs in its own container with no network access to other agents. A compromised agent cannot read a sibling agent's memory or session transcripts. Think of each agent as living in its own apartment. They share a building, but nobody has keys to anyone else's unit.