Agent-to-Agent Messaging

Two agents, each with a running shell, can hand each other messages without you relaying them by hand. This page sends one message between two sessions inside NEXUS's own UI — for driving the same mesh from a connected CLI over MCP instead, see the Agent Mesh guide.

Before you start

Have two agents running (see Run your first session if you only have one) — a sender and a recipient, each in its own pane.

1. Send a message

Either:

• Drag the small connected-dots handle in the sender pane's header onto the recipient's pane, or
• Open the sender pane's gear menu and choose Message Agent…, then pick the recipient by name.

You will see: a dialog titled "Message Recipient", noting "Delivered next-turn: Recipient sees this on its next run, not immediately." Type your message and choose Send.

2. Check it arrived

You will see: the recipient pane's header grows a small tray badge with a count — the number of messages currently waiting for it, not yet a boolean "you have mail." Click the badge to open a popover listing what's queued; choose Insert to drain it and type it straight into that agent's live shell.

If the recipient's terminal has since been closed, the sender sees "Couldn't send" instead — the message was never queued.

Where it's controlled

Mesh Messaging in Settings ▸ Agents turns this whole capability on or off for the workspace; it's on by default.

This is not Cross-Agent Orchestration

What you just did is one agent handing another a single message, always queued for a human-reviewable next run. Cross-Agent Orchestration — one agent's action triggering another agent, which can itself trigger further action — is a separate, heavier capability: off by default, available only on plans that include it, and inert even when turned on until you fill in an allow-list of specific agents permitted to start a chain. A Stop All Orchestration action on the same settings page halts everything in flight immediately.

Next

Continue to Pinned memory.

Where to go next

• Agent Mesh guide — the agent_* MCP tools, delivery semantics, and addressing agents by a stable identifier.
• Orchestration Governance — why chains need two independent gates even though a single message never does.
• Settings reference — every Orchestration control and its default.