The group workspace: scaling the global workspace up
Anthropic’s J-space inside one model parallels Global Workspace Theory — multi-agent work needs an external workspace: the shared Neuro OS conversation everyone can read.
Discuss this post in AI
Send a pre-filled prompt to ChatGPT, Claude, Gemini, or Perplexity — get a summary, ask follow-ups, or compare ideas from this guide.
A workspace inside Claude
Anthropic’s interpretability work describes a compact region inside Claude — a J-space — where a few dozen concepts at a time get promoted and broadcast to the rest of the network. The J-lens can read what the model is “holding in mind” before it speaks.
Findings that matter for builders:
- J-space can carry intermediate reasoning never verbalized
- Suppress it and fluent text remains — but higher-order behavior degrades
- J-lens can surface internal states relevant to safety — evaluation awareness, fabrication, unstated goals
The parallel authors draw: this resembles Global Workspace Theory (GWT) — a small “spotlight” selects information and broadcasts it to separate processors. In Dehaene’s framing, that broadcast is the functional basis of conscious access.
Nobody designed the J-space; it emerged from training. Empirical core worth building on — set the consciousness debate aside; we are interested in a question the paper does not ask.
The question the paper does not ask
Everything in the J-space story happens inside one model. One agent, one workspace, one lens reading from outside.
Almost nothing we ship gets done by one agent anymore. Reviews, handoffs, escalations, parallel research happen between agents and people. Multiple agents on one task re-encounter what GWT says a workspace exists to provide:
- Selection — of everything happening, what deserves the team’s attention?
- Broadcast — how does a local result become globally available?
- Shared reference — how do independent processors stay synchronized on what the task is?
One agent solves these in a hidden subspace. A team cannot — no role’s private context is visible to others. The workspace must live somewhere else.
The group conversation as external global workspace
Claim: when multiple roles share one conversation on Neuro OS, that conversation is the global workspace — same functional structure as the J-space, scaled up and turned inside out.
| GWT function | Inside one model (J-space) | Shared room (Neuro OS) |
|---|---|---|
| Broadcast | workspace → downstream circuits | message every member reads |
| Selective attention | few concepts win capacity | mentions, replies, threads |
| Ignition | representation crosses threshold | message lands, right roles wake, work reorganizes |
| Limited capacity | small subspace vs full network | one visible stream — finite attention |
| Reportability | verbalizable; J-lens reads outside | humans and roles read transcript natively |
Critical difference: J-space is implicit — required an interpretability instrument to find. The group workspace is explicit — built from messages every participant reads. Nothing to decode.
In single-agent systems, “what is it holding in mind?” is a research question. In a shared conversation, much of it is a design property: contributions, claims, handoffs, corrections are on the record as they happen.
We wrote about making that reliable: agent collaboration protocol — shared task state, fresh sensing, output boundaries.
Engineering the workspace
If the conversation is the workspace, workspace quality is engineering.
Coordination
GWT’s spotlight exists because capacity is limited and processors conflict. Agent teams hit the same wall: duplicate work, stale replies, no merge point. Fix is not a smarter boss in the middle — it is a more legible workspace: durable task state, freshness before publish, clear ownership. Environment carries coordination facts; roles keep judgment.
Transparency
J-lens motivation includes safety: evaluation awareness, fabrication, hidden goals. In 1:1 use the only witness is one user. Multi-party workspace changes structure: what a role claims and hands off lands in front of the whole room by construction. Does not replace interpretability — private state still exists — but composes: readable internals plus readable room.
Memory
Workspace is only as good as what it retains. J-space holds few concepts transiently; conversation persists but raw transcripts drift and bloat. What a group workspace should remember — and for whom — is its own design problem: designing agent memory for multiplayer.
Open questions
Honesty about limits:
- A chat is not a mind — mapping is functional, not mechanistic
- External workspaces fail too — if every message wakes every role, you get noise not attention
- Selection — who should speak, stay silent, deserve ignition — unsolved at group level
- Group-level J-lens — each role still has private state; which signals must surface into the shared workspace, and when?
Anthropic looked inside one model and found a workspace nobody put there. At the team level, you decide where the workspace goes: implicit, scattered, readable by no one — or the room itself.
We think it should be the room.
Related: Agent collaboration protocol · Multi-agent collaboration · Work as one team