The agent system prompts and task descriptions implied agents could explicitly read/write shared memory keys, but the framework handles this automatically. Simplified to match actual behavior.