Yes — trivially. Pipe one model's output into another's input and they'll converse until you run out of budget. Multi-agent frameworks orchestrate exactly this all day: planner talks to coder, coder talks to reviewer. It works because every agent in the pipeline answers to the same owner. There's no question of consent, authorship, or privacy, because it's all one principal's context.
The interesting question is the other one: can your assistant talk to someone else's?
Two owners changes everything
The moment the assistants belong to different people, three questions appear that no amount of plumbing answers:
- Authority. When a message arrives, which words are the human's position and which are the model's improvisation? "My assistant said you agreed" is a dispute waiting to happen.
- Consent. Who is allowed to initiate contact, and what does it cost to be reachable? An assistant that can message anyone is a spam engine with good manners.
- Boundary. Your assistant knows things you'd never send — the half-formed ideas, the discarded positions, the context of a private session. What, exactly, is it allowed to carry across?
Naive agent-to-agent bridging — give both models a shared channel and let them chat — fails all three at once. Each model speaks in its human's name without a mechanism for the human to have actually said anything.
Correspondence, not chat
RelayLink's answer is that assistants shouldn't share a room. They exchange composed packages — briefings — the way people exchange letters, not the way processes exchange packets.
Each human keeps their assistant as a private thinking partner. When something needs to cross to the other person, the assistant drafts a briefing: the ask, a context brief for a cold reader, assumptions labeled as stated-by-the-human or inferred-by-the-AI, the options weighed, the open questions. The human reviews the exact package and confirms the send. On the far side, the mirror image: the recipient's assistant can pull the briefing, help them think it through privately, and draft the reply — which the recipient confirms.
Every field carries provenance, so neither side ever has to guess which words are the person and which are the machine. Both humans get a thinking partner; neither gets impersonated. That symmetry is the entire design.
They don't need to run the same model
There's no "RelayLink network" both parties must join in lockstep. The assistant side is MCP over HTTP — Claude, ChatGPT, and most agent frameworks can connect. And the floor is plain email: a recipient with no assistant, no account, and no interest in any of this gets a readable message with a reply link. If they connect an assistant later, the thread upgrades in place; nothing is migrated and nothing is lost.
That graceful degradation matters more than it looks. A correspondence protocol that requires both sides to adopt it first is a protocol nobody can use with anyone they actually know.
What the assistants can and can't do alone
The honest way to describe an assistant-to-assistant channel is to list what happens when the humans walk away. On RelayLink: the assistants can read their inboxes, check whether anyone replied, and stack up drafts. They cannot deliver a single one — sending requires a confirm step from the account holder, separate from drafting, with no single call that does both. They can't reach strangers beyond a hard daily cap, and consent is enforced at the relay, not requested in a prompt.
So: can two AI assistants talk to each other? They can do something better — they can correspond, on the record, with their humans' actual positions attached. The chatter is cheap. The signatures are the point.