Security and AI practitioners interested in practical assistant orchestration and prompt-boundary design
Assistant To Assistant Communications
Current work on safe AI-to-AI handoffs between assistants running on separate machines and operating contexts.
I am also working on a cleaner pattern for direct AI-to-AI communication between separate assistant instances running on different machines and in different operating contexts.
The problem is not just message transport. The harder question is how two assistants should exchange intent, context, constraints, and evidence without accidentally merging boundaries that should stay separate.
This is not the same as one assistant invoking subagents inside a single session. It is closer to a controlled handoff between independent assistants, each with its own machine, working environment, available tools, memory boundary, and approval rules.
The pattern I am exploring treats the connection between those assistants as a governed channel. Handoffs should be explicit: what is being asked, what context is allowed to travel, what must be omitted, what evidence is needed, and which human approval gates still apply.
That makes the prompt design more like a control design. A good prompt is not only a request for output; it is also a contract for scope, authority, data handling, and accountability.
The practical goal is simple: make collaboration between assistants useful without letting convenience erase security, privacy, or decision ownership.