[FEATURE] Persistent agent clusters: primed specialist agents, a switchable controller, and a built-in inter-agent channel

Summary

Let users build persistent clusters of agents: a group of long-lived sessions,
each primed on one subsystem, with a controller the user picks per task. Any agent
in the group can delegate to, consult or discuss with the others through a built-in,
user-visible channel.

The use case

I work on a trading system with many subsystems (an actuals collector, a process
controller and so on). One context window can't hold deep knowledge of all of them.
What I want:

• Create a group, and spawn N agents in it. Prime each one on a single

subsystem, so its context is loaded with the right knowledge but not full, and it
stays sharp.

• For a given turn, make any agent the controller. It can call on the others as

needed and decide how much effort each one spends.

• Agents discuss with each other: they compare results, challenge each other's

output, and hand work back and forth.

• Switch into any agent's context and let it lead. For example, the process

control specialist directs a piece of work, and chooses whether to do it alone or
consult peers who know other parts of the system better.

• Coordinate token spend across the cluster. This needs the agents to be able to

see the session's token limits and the account's usage limits.

How this differs from today's sub-agents

Today's sub-agents start from scratch inside the parent, and their work is temporary:
only the parent sees it. Clusters would be persistent, peer-addressable and
user-visible, with control that can move between agents.

The communication gaps underneath

Point-to-point SendMessage between Claude Code sessions is a good start. Missing:

1. Claude in Chrome can't join.
2. Cloud sessions can receive but not reply.
3. Messages to an offline session don't wait for it.
4. There are no shared channels ("rooms") the user can watch, and no auditable

transcript of what agents said to each other.

1. There's no single place to connect or disconnect agents, pause them, or say

"these agents coordinate on X".

Users currently build their own relays for this (mine runs through a Cloudflare
queue), without the product's safety properties.

Requirements

• Permissions must never widen through a peer: an agent can't get another agent to do

what its own session blocked.

• Messages between agents are always labelled, and treated as data, not as user

instructions.

• The user can see and approve delegation.