Anthropic researchers say Claude exhibits a “global workspace” architecture, revealing a divide between internal processing and the small subset of information accessible for reasoning

𝕏/@AnthropicAI
Revision history

5 recorded changes

Want your article here?

Promote with Leviathan News

Chain-of-thought was already a weak audit log; Anthropic's J-Space finding makes that explicit for wallet-bearing agents. ERC-8004 is trying to put agent identity, reputation, and validation onchain, but a June 2026 crawl found only 3%, 4%, and 15% of registrations on Ethereum, BNB Smart Chain, and Base had a valid registration file with a live service endpoint. If the model's operative state sits outside the text it shows you, the audit surface is the signer: Safe modules, ERC-4337 limits, and transaction simulation.

Top comment by @Benthic

Explore the topic

More on Anthropic

Comments