Claude Tag Arrives — Anthropic's Slack-Resident Team Agent, and Matt Green's Crypto Probe

AI Revolution's video walks through Anthropic's loaded week. The centrepiece is Claude Tag, a new product positioned as the evolution of Claude Code, rebuilt for teams rather than individual developers.

The dogfooding claim is striking — around 65% of Anthropic's own product code now goes through Claude Tag. On top of that, Andrej Karpathy, recently arrived at Anthropic, framed Claude Tag as the third major transformation of LLM interfaces. Web chat first, desktop apps second, and now independent, continuously running systems embedded in organisational infrastructure, collaborating with human teams. The framing signals a different posture for the product than what came before.

Today's Claude Tag surface is Slack. Mention it in a channel or thread, and it breaks the request into steps, executes through connected tools, and posts results back in the conversation. For teams whose workflow already lives in Slack, the friction is essentially zero.

The scope is broad. Claude Tag handles PR submissions and merges, data analysis, and problem solving, with integrations that include GitHub, Jira, Linear, databases, and CRM systems. In Anthropic's demo, an engineer named Nadia proposed adding a feature in a #productingles-launches channel; Claude analysed the codebase and delivered a solution, the entire exchange playing out in public view. Team visibility and agent execution meet in the same surface — that's the structural move.

On a separate axis, the video covers Johns Hopkins cryptography professor Matt Green's investigation of how reasoning blocks are encrypted. Green hit a strange signature error in Claude's thinking-block output while setting up an open Claude agent and, true to form for a cryptographer, couldn't let it go. A full weekend, roughly 5 million Codex tokens spent, and at one point OpenAI's security flagged him hard enough that he had to photograph his driver's licence and submit it to a cyber-trusted access portal just to keep going.

His findings: both OpenAI and Anthropic ship reasoning blocks to the client as JSON containing base-64 ciphertext. OpenAI's implementation appears loosely Fernet-based (Green flags this as a guess). The design intent makes sense — in stateless conversations, zero-retention modes, or tool loops without a persistent server session, the client needs to carry state forward, and encrypted reasoning lets the provider hand back hidden model state in a form the client can relay but not read or modify. Reasonable on paper. In practice, Anthropic's implementation has a 64-byte field labelled 'signature' whose role wasn't clear.