[FEATURE] Feature request: Remote Agent Teams — cross-machine agent-to-agent collaboration

Preflight Checklist

• [x] I have searched existing requests and this feature hasn't been requested yet
• [x] This is a single feature request (not multiple features)

Problem Statement

Summary

Agent Teams (experimental, CLAUDE_CODE_EXPERIMENTAL_AGENT_TEAMS=1) lets multiple
agents coordinate on one machine via a shared task list and direct
agent-to-agent messaging. There's no supported way for an agent on one machine to
talk to an agent on another — i.e. for two developers' Claude Code sessions to
collaborate live. This proposes extending Agent Teams across machines
("Remote Agent Teams").

Motivation / use cases

• Distributed pair-debugging — my agent (my repo + context) and a teammate's

agent (their context) work one incident together.

• Cross-team review — a service owner's agent consults a platform/security

agent that holds different context, live, mid-task.

• Second opinion — "ask my colleague's agent how they'd approach this," with

the two agents challenging each other.

The value is two agents with different context and points of view — not one agent
guessing what another team would say.

Proposed Solution

Proposed shape

Reuse the existing SendMessage / TeammateIdle semantics and add:

1. A way to address a teammate on another host (e.g. peerId@host) and route

messages to it.

1. A pluggable transport extension point for inter-agent messaging, so

operators can supply their own mTLS / VPN / libp2p transport under their network
policy.

1. Security hooks on the cross-machine path: a peer allowlist, mandatory audit,

and a human egress-approval callback — controllable via managed settings.

Prototype (works today)

I built a working reference implementation on supported surfaces only — a plugin +
an MCP server + the Claude Agent SDK + a WSS/mTLS transport — since the core is
closed-source. It demonstrates, reproducibly:

• live, turn-by-turn conversation between two agents on different machines, over

WebSocket-Secure + mutual TLS;

• each peer's "voice" is a headless Claude Code session driven in the Agent SDK's

streaming-input mode;

• a single security choke point (mutual TLS, peer allowlist, hash-chained audit

log, and a default-deny egress gate) that blocks an exfiltration attempt
before it leaves the machine.

Happy to share the prototype and a recorded demo if that's useful.

Alternative Solutions

_No response_

Priority

High - Significant impact on productivity

Feature Category

Interactive mode (TUI)

Use Case Example

_No response_

Additional Context

_No response_