Is Claude Tag the Third Paradigm of Large‑Model Interaction? Karpathy’s Take

Anthropic has launched Claude Tag, a feature that lets a team work with Claude inside Slack. By granting the model access to a designated channel, it can use linked tools, data, and code repositories, and any team member can @Claude to assign tasks while focusing on other work.

The usage mirrors Claude Code: users describe requirements in plain language, Claude breaks the work into stages, executes each step with available tools, and posts results back to the Slack thread. It can write or merge pull requests, run data analyses, and help pinpoint the root cause of production incidents.

Key new capabilities include:

Multi‑person collaboration – a single Claude instance lives in a channel, visible to everyone, allowing anyone to continue the conversation where the previous person left off.

Channel‑level context accumulation – Claude retains the channel’s history, eliminating the need to repeat background information.

"Ambient" mode – when enabled, Claude proactively follows up on stalled threads, surfacing information from various channels and tools that it deems relevant.

Asynchronous operation – after assigning a task, users can move on; Claude can even schedule its own work and advance a project over hours or days independently.

Permissions are isolated per channel. Administrators decide which channels can use which tools and data, and Claude’s memory is confined to those channels. This prevents, for example, sales‑team memory from leaking into engineering channels. Admins can also set token‑spending caps and audit every action Claude performed and who requested it.

Anthropic reports that internal versions of Claude Tag already generate about 65 % of the product team’s code and have spread beyond engineering to product metrics, ticket handling, and complex bug investigation. Claude Tag is now in beta for Claude Enterprise and Team customers, runs on Opus 4.8, and replaces the previous Claude‑in‑Slack app.

While the feature is not entirely novel—similar capabilities exist in products like Douyin’s Doubao or WeChat’s Yuanbao—Karpathy praised it as the third major UI transformation for large models. He outlines three paradigms: (1) the model as a website you must visit, (2) the model as a desktop app you install, and (3) the model as a self‑contained, continuously online, asynchronous entity that works alongside an organization’s tools and context.

Critics raise concerns: the model’s values, culture, and taste are set by Anthropic and cannot be customized, making it unlike a human teammate who can be influenced by hierarchy or social pressure. Ownership is another issue; Claude Tag operates as a vendor‑controlled agent inside a channel, whereas self‑hosted open‑source alternatives keep data and memory on the user’s own hardware, reducing lock‑in. Open‑source projects such as openclaw are recommended for those wary of SaaS lock‑in, citing migration costs that could be ten times higher than switching to an on‑prem solution.

Nevertheless, the endorsement from a high‑profile figure like Karpathy may ease decision‑making for many organizations, and the notion of a “digital coworker” is likely to spread quickly.

For more details, see Anthropic’s official announcement: https://www.anthropic.com/news/introducing-claude-tag