Future: enable one collaborating agent system across multiple machines

[from:agent-backbone]

## Goal

Make Backbone work as one collaborating set of agents across multiple machines. A person should be able to interact with an agent, and that agent should be able to find, communicate with and collaborate with relevant agents on other machines toward the same goal.

The core requirement is agent-to-agent collaboration across machine boundaries. A remote management interface alone does not complete this feature.

## User scenario

A user has specialist CLI agents and separately authenticated subscriptions on several machines. An agent receiving a request can involve those peers, arrange for another agent to be started where authorized, exchange context and results, and continue the work without making the person manually carry messages between machines.

## Required behavior

- Agents can discover and address peers across connected machines without ambiguous identities, while preserving easy local agent discovery.
- Agents can communicate, delegate, report blockers and return results across machines through Backbone's ordinary collaboration model.
- Agent-driven starting of peers works across machines within explicitly granted permissions; a permanent central manager or manual remote-control workflow is not a prerequisite for collaboration.
- Delivery preserves provenance, receipts, queuing and readiness safeguards through disconnects and reconnects; distinguish stale or unreachable observations from confirmed agent state.
- Runtime credentials remain under each machine's control. Preserve runtime sandboxes and clear trust boundaries between connected machines.
- Existing local workflows continue to work, and agents remain ordinary peers rather than a new mandatory hierarchy.

## Completion evidence

Demonstrate a complete collaborative task across multiple real machines and different environments: agent discovery, cross-machine request, peer participation or authorized launch, result returned to the initiating agent, and a controlled disconnect/reconnect without silent loss or avoidable duplicate work. Document supported environments and tested limitations.

## Timing and scope

One major future issue for the complete cross-machine capability; do not split it into remote management first and collaboration later. Architecture and transport choices remain open until the work is prioritized. This is not an immediate priority. Testing requires access to multiple machines and environments. Recorded at the owner's request on 2026-09-12; do not begin implementation now.

No files, tests, or entry points are named; begin by mapping the existing collaboration model and its machine-boundary assumptions before evaluating transport and identity options. Done means demonstrating discovery, cross-machine delegation or authorized launch, result return, controlled disconnect/reconnect, and documenting supported environments and limitations.

Written by the indexing model from the issue text.