What is Anthropic's global workspace in Claude? The J-space, explained

Anthropic just published research called A global workspace in language models, and it is one of those findings that quietly reframes how you should think about the agent in your terminal. The short version: Claude has two ways of thinking, and almost every frustration you have had with an AI coding agent comes from it using the wrong one.

What Anthropic found

Borrowing from a neuroscience idea, the researchers looked for a small shared channel inside the model, a place where information becomes available to the rest of the system for deliberate use. They found one. They call it the J-space, and it behaves like a workspace: the model can report what is in it, deliberately control it, and use it to carry multi-step reasoning. Crucially, most automatic processing, grammar, fluent phrasing, simple recall, never touches it. It emerged on its own during training. Nobody built it.

Two examples make it concrete. Ask Claude to compute a small sum while copying an unrelated sentence, and the intermediate numbers show up only in this internal space, never in the output. And when Claude reads buggy code, the concept of an error lights up in the workspace before it writes a single word of response. The model often notices the problem internally before it acts.

One honest caveat, because it matters: this is about what researchers call access consciousness, information the model can report and deliberately use. It is not a claim that Claude has subjective experience or feelings. The interesting part for us is not whether the model is conscious. It is that deliberate reasoning lives in a different place than automatic output.

Your agent has two modes, and the default is the wrong one

Put that next to how a coding agent misbehaves. Out of the box, an agent runs on the automatic setting: it produces fluent output fast. It guesses what you meant instead of stopping to ask. It writes the most complete-looking version instead of the smallest one. It says the job is done the moment the code exists, before anything ran. That is the autopilot, doing what autopilot does, generating confidently.

But the research shows the deliberate workspace was there the whole time. The model that types "this should work" over broken code is, in some cases, the same model whose workspace already flagged the error. The capability for careful, checked reasoning is not missing. It is just not the default path, and nothing forced it to switch on.

Why "ask first" and "prove it" actually work

This is the part that turns a research paper into something you can use tonight. The habits that reliably improve AI coding output are not magic prompts. Every one of them is a way of forcing the deliberate workspace to engage instead of letting the autopilot run.

• Planning before coding makes the model reason through the whole change in the open before touching a file. That is deliberate reasoning by definition.
• Asking before building interrupts the guess. A guess is autopilot. A question is the workspace admitting it needs information it does not have.
• Proving a change runs before calling it done forces the model to surface what its workspace may have already noticed, instead of papering over it with a confident sentence.

This connects back to something we wrote about earlier, that an archetype is a value function, not a skill. The deliberate workspace is where a value function can actually take hold. You are not making the model smarter. You are changing which mode it does the work in.

Building the deliberate mode in by default

That is the entire premise of kitstarter. The default agent runs on autopilot, so kitstarter.dev ships a CLAUDE.md and a set of commands that make Claude Code, Codex, and Antigravity plan first, ask before they build, and prove their work runs, the exact behaviors that pull a task into the deliberate workspace instead of the autopilot. Anthropic found the careful mind was in there all along. The job is just to make your agent use it.

Common questions

What is the global workspace in Claude? A small internal space (the J-space) that acts like a shared channel for deliberate reasoning, distinct from automatic processing. It emerged on its own during training.

Does this mean Claude is conscious? No. It points to access consciousness in a functional sense, information the model can report and use, not subjective experience.

What does it mean for using Claude Code? Your agent has a fast autopilot and a slower deliberate mode. Bad output comes from the autopilot. Plan, ask first, and prove it are how you engage the deliberate mode.