Anthropic Claude Tag: An Always-On AI Teammate in Slack

Anthropic has turned Claude into a standing member of your Slack workspace. The new Anthropic Claude Tag, launched in beta on June 23, 2026, lets anyone summon the company's AI by typing @Claude in a channel and handing it a task in plain language. According to TechCrunch, the feature is positioned as an always-on teammate rather than a chatbot you call up and dismiss, and it is built to accumulate knowledge of how a company works the longer it sits in a channel.

What the Anthropic Claude Tag actually does

The mechanics are straightforward on the surface. A user tags @Claude in any channel and describes what they want in ordinary language. The Decoder reported that Claude then breaks the request into steps, executes them using whatever tools it has been connected to, and reports the results back inside a Slack thread so the work is visible to everyone. That last point is more important than it sounds. Instead of a private exchange between one person and a bot, the model's reasoning and output sit in the open channel where colleagues can read along.

This is what Anthropic calls multiplayer. Each channel has a single shared Claude instance, and any member can pick up a conversation someone else started, see the steps the AI took, and keep going. The design treats the AI less like a personal assistant and more like a coworker whose desk everyone can see. The Decoder noted that private direct messages are still available for individually configured tools, so a person can keep certain work to themselves, but the default posture is shared and transparent.

Beyond the synchronous request-and-reply pattern, the feature adds two capabilities that change its character. The first is a proactive mode. When enabled, Claude surfaces relevant information and follows up on stalled threads without being asked, according to The Decoder. The write-up from Latent Space described this as Claude responding to channels without being tagged, reviewing messages to decide when a response is needed, and watching for threshold events that might trigger a fix or a monitoring action. The second is persistence. The model builds context from a channel's history and can schedule its own tasks across hours or days, which means it can wait on a blocking dependency, then act when the condition is met.

From solo tool to multiplayer agent

Latent Space framed the launch as a deliberate shift in how Anthropic wants Claude to operate, moving from a solo, synchronous tool toward a multiplayer agent that works asynchronously and acts on its own. The contrast is with Claude Code, the company's coding assistant that a single developer drives in a single session. Claude Tag keeps the same underlying capability but rehouses it inside a shared workspace where delegation is the point.

The practical features reflect that goal. Latent Space reported that a user can tag Claude in and delegate a task while focusing on other work, and can also tag in coworkers who own related code. The agent supports what the write-up called stacked prompts rather than just stacked diffs, and it can hook into git webhooks that wait for blocking dependencies over very long periods, measured in days. It can also summarize a long thread into a document with action items. The throughline is that the work does not have to happen in one sitting, and it does not have to involve only one person.

"Tag Claude in and delegate tasks to it while you focus on other work."Latent Space, on Claude Tag's multiplayer model

The 65 percent claim, and what it leaves out

The headline statistic from the launch is that Claude Tag already generates 65 percent of the code on Anthropic's own product team, a figure the company supplied and The Decoder reported. Anthropic says adoption has spread beyond engineering to people handling product metrics and support tickets, plus bug diagnosis. The implication is that the tool has graduated from experiment to daily dependency inside the company that built it, and Latent Space noted that Anthropic's Claude Code team had used it internally all year before the public beta.

That number deserves a careful read. As Latent Space pointed out, the claim appears in two slightly different forms, writes 65 percent of the product team's code and merges 65 percent of product pull requests, and neither version comes with a denominator, a timeframe, or a clear definition of what counts as authorship versus a merge. A line suggested by an AI and edited by a human is very different from a fully AI-written pull request reviewed by a person, which differs again from a human-written change that Claude merely merged. The figure of 65 percent could describe any of them. First-party productivity figures from an AI vendor about its own AI are useful as a signal of internal confidence, not as a benchmark you can port to another company. The honest summary is that Anthropic leans heavily on the tool and wants customers to know it.

How permissions and scoping work

The most consequential design decisions sit in the access model, because an AI that reads a company's Slack is also an AI that could read the wrong channel. The Decoder reported that administrators control which tools and data Claude can reach on a per-channel basis, and that admins can create separate AI identities with isolated memories. TechCrunch added that this scope control prevents overlap, giving the example that a Claude configured for legal work cannot influence engineering channels. Claude can gather facts from other channels automatically, but only where permission has been granted.

Auditability is part of the pitch. The Decoder noted that full action logs track every operation and who triggered it, which matters for enterprises that need to answer who asked the AI to do what. Latent Space, writing from a more technical angle, flagged the parts that remain undisclosed: the exact granularity of access control, how secrets and credentials are handled, how data retention and memory are scoped, and the full security model. Those gaps are not unusual for a beta announcement, but they are the questions a security team will ask before turning the feature on across a workspace. An agent that is proactive and persistent is, by definition, doing things when no one is watching, which raises the bar for logging and scope enforcement.

The isolated-identity design is the part most likely to reassure cautious buyers. By letting an administrator stand up a separate Claude per function, each with its own memory and its own connected tools, Anthropic is acknowledging that a single all-knowing agent across an entire company is a liability rather than a feature. A legal team's Claude should not be able to read source code, and an engineering Claude should not be reasoning over confidential HR threads. Whether the boundaries hold under real use, especially when the agent is allowed to pull facts from other channels with permission, is the kind of thing that only becomes clear once large organizations stress-test it. The promise is clean separation; the proof will be in the incident reports that do or do not appear over the coming months.

The real strategy is organizational context

TechCrunch was direct about the bigger game. It described Claude Tag as a strategic play to capture organizational context, institutional knowledge, and enterprise workflows. The productivity story is the surface; the deeper value is that the longer Claude sits in a company's channels, the more it learns about how that company actually operates. TechCrunch put it plainly, noting that as Claude follows along with its channel, it learns ever more about the work.

That is a defensible position to occupy. Generic model quality is converging across vendors, and the thing that becomes hard to copy is a model that has absorbed how a specific organization works: its history, its jargon, the decisions it has made and the unwritten rules it follows. An AI that knows where the bodies are buried in a codebase, which customer complaints recur, and how a given team makes decisions is far stickier than one that answers questions from a blank slate. By embedding in Slack, where a great deal of a modern company's institutional memory already lives, Anthropic is positioning Claude to become the layer that understands the business, not just a tool that completes tasks.

A crowded race for the enterprise intelligence layer

Anthropic is not alone in chasing this. TechCrunch named Microsoft, with its Graph and Copilot stack, alongside Snowflake and Databricks, as well as Glean, among the competitors pursuing a similar enterprise intelligence layer. Each is trying to become the system that sits on top of a company's data and turns it into answers and actions. Microsoft has the advantage of owning the productivity suite many enterprises already run, while Glean built its business specifically on enterprise search across a company's apps.

Anthropic's angle is to enter through the place where work conversation happens rather than through documents or a search box. Slack is where decisions get made and questions get asked, often more candidly than in formal docs, and where context gets shared as a matter of course. Embedding an agent there is a bet that the conversation layer is the right beachhead. It is also a reminder that for Anthropic, which sells through partners and API access more than through a consumer app, enterprise penetration is the commercial heart of the business. Claude Tag is a direct route into the daily workflow of paying customers.

Availability and the migration path

Claude Tag is in beta for Claude Enterprise and Claude Team customers, according to The Decoder, and runs on Opus 4.8, Anthropic's frontier model as of this launch. None of the coverage disclosed specific pricing for the feature. The Decoder also reported a concrete operational detail for existing users: Claude Tag replaces the older Claude in Slack app, and Anthropic is giving customers a 30-day window to migrate. Teams that have wired the old app into their channels will need to plan that transition rather than assume continuity.

Slack is the first and, for now, only platform, with Latent Space noting that future platforms were left unannounced. That is a sensible place to start given how much enterprise communication runs through Slack, but it also means the feature's reach is tied to a single tool for the moment. Companies standardized on Microsoft Teams or other platforms will have to wait, or rely on Anthropic's other integration paths.

Why proactive agents are the industry's next step

Claude Tag fits a broader pattern that has been building for much of 2026: the move from agents you summon to agents that run in the background. The first wave of AI assistants was reactive. You opened a window, asked a question, and got an answer. The capability was impressive, but the interaction model put all the burden on the human to know when to ask. A proactive agent inverts that. It watches a stream of work and decides for itself when to step in, which removes the friction of remembering to delegate but introduces a new problem: an agent that acts on its own can act wrongly on its own.

That tension is the central design challenge for everything in this category. Too timid, and the proactive mode adds no value over a plain chatbot. Too eager, and it floods channels with unwanted suggestions or takes actions on context it misread. Anthropic's decision to make the work visible in the open channel is partly an answer to this. If the agent's reasoning and output sit where colleagues can see them, mistakes are caught faster than they would be in a private exchange. Transparency becomes a safety mechanism as much as a collaboration feature. The persistence angle compounds both the promise and the risk, because an agent that holds context for days can pick up exactly where a project left off, but it can also carry a stale assumption forward long after the situation has changed.

For Anthropic, shipping a polished version of this into a tool as widely used as Slack is a way to set the terms of the conversation before competitors do. The company has historically emphasized safety and interpretability in its public messaging, and a launch that foregrounds admin controls and scoped identities, along with full audit logs, is consistent with that posture. Whether enterprises read those controls as sufficient, or as the opening bid in a longer negotiation over trust, will shape how fast Claude Tag moves from beta into broad deployment.

What to watch as Claude Tag rolls out

The launch lands at a moment when background and proactive agents are the clear direction of travel across the industry, and Claude Tag is one of the more fully realized versions of that idea to ship into a mainstream workplace tool. The capabilities Anthropic describes, an agent that acts without being asked, that remembers across days while staying visible to the whole team, are genuinely a step beyond the reactive Slackbots that came before.

The open questions are the ones that always trail an ambitious enterprise launch. How tight is the scoping in practice, and can it be trusted to keep legal context out of engineering channels and credentials out of the wrong hands? How will the proactive mode behave at scale without becoming noise, and what happens when a persistent agent acts on stale context? And how much of the 65 percent internal-adoption story will translate to companies that did not build the tool and do not staff teams of people fluent in working alongside it? The beta period is where those answers will come from. For now, Anthropic has made a clear statement about where it thinks AI belongs: not in a separate window you visit, but inside the conversations where work already happens.

The competitive stakes give the launch added weight. Model quality alone is no longer a durable moat, since rival labs ship capable systems within months of each other. What endures is distribution and embedded context, and Claude Tag attacks both at once by planting Anthropic's model in the daily workflow of paying enterprise customers and letting it accumulate knowledge that would be costly for a competitor to replicate. If the beta holds up and the trust questions get satisfying answers, this could become one of the stickier products in Anthropic's lineup. If the proactive behavior proves noisy or the scoping leaks, it will serve as a cautionary case study for everyone else building background agents. Either outcome will teach the industry something it needs to know.