AI Just Joined the Team. Here's the Work That Makes That Real.

The interface changed. That's the smaller story.

On June 23, 2026, Anthropic shipped something that sounds like science fiction and is now just a beta you can turn on: Claude Tag. You add Claude to a Slack channel and it joins as a team member — a shared, visible participant anyone can @-mention to hand off work. It remembers the channel's history. In "ambient mode" it can follow up on a stalled thread without being asked. It is, functionally, a coworker who happens to be a model.

The launch got the coverage you'd expect. But the most useful thing anyone said about it didn't come from Anthropic. It came from Andrej Karpathy, who framed the shift at the level that actually matters:

That's the right frame, and it's worth pausing on who said it. This isn't Anthropic's marketing — Anthropic described Claude Tag modestly, as "an evolution of Claude Code." The paradigm claim is an outside observer's read on where things are going, which is exactly what makes it credible rather than promotional. Three interfaces, each a bigger leap than the last: the tab you open, the app you install, and now the colleague you delegate to.

Here's why this piece isn't really about Claude Tag, though. Because Karpathy, in the very next breath, named the thing that decides whether any of it works for your company — and almost everyone skipped past it.

The part of the quote everyone skipped

The clause that matters comes right after the paradigm line. Karpathy said the teammate model is "significantly more 'inline' with all the other human activity org-wide" — and then, in a parenthetical most readers glossed over, he named the price of admission: "Once you do all of the under the hood engineering work to make this 'just work' (e.g. across tools, integrations, compute environments, memory, security, etc.)."

Read that parenthetical again, because it's the whole article. Once you do all of the under-the-hood engineering work. Tools. Integrations. Compute. Memory. Security. The person who coined the "AI teammate is a new paradigm" framing is telling you, in the same sentence, that the paradigm sits on top of a mountain of unglamorous infrastructure — and that "just works" is a state you engineer your way into, not a feature you switch on.

The person who framed the AI teammate as a new paradigm — Andrej Karpathy — named its precondition in the same breath: "all of the under-the-hood engineering work to make this 'just work'… tools, integrations, compute environments, memory, security." The teammate is the interface. The engineering underneath is what makes it real. Most of the excitement is about the first. Almost all of the difficulty is in the second.

This is the distinction we've been making for a year, now handed to us by the field's most-cited voice: the difference between using AI and building on it as a system. A teammate who can be trusted with real work needs to reach your tools, hold your context, remember what happened, and be contained by your security boundaries. None of that is the teammate. It's the house the teammate moves into. And whether that house is built is the entire difference between "AI joined our team" and "we added a bot to a channel and nothing changed."

What Claude Tag actually is — and what's underneath it

Let's be concrete, because the specifics make the point better than the abstraction. Claude Tag isn't a private chatbot bolted onto Slack. It's a shared, "multiplayer" team member: one Claude per channel, visible to everyone, whose work happens in the open thread rather than in a hidden 1:1. Admins decide which channels it's in and which tools and data sources it can reach. You @-tag it to delegate; it breaks the task into stages and posts as it goes. It builds memory from the channel's history. Right now it's a beta, available to Enterprise and Team customers, and it lives in Slack first — Anthropic has said it intends to bring the same taggable teammate to the other places teams work.

Now look at what Anthropic had to build under that simple experience, because it's a map of what "AI teammate" actually requires. Each Claude identity is scoped per channel — a Claude configured for a legal channel can't carry its memory into an engineering one. Each thread runs in its own sandbox that holds no stored credentials; access is injected at request time and revoked after. Outbound network access is default-deny. That is not product polish. That is the security and identity architecture that has to exist before you'd let any new participant — human or model — into a room where real company information lives.

And here's the number everyone will quote at you, with the caveat that makes it useful instead of misleading: Anthropic says 65% of its own product team's code is now created by its internal version of this system. That's real — and it's a statement about Anthropic, a company with flawless internal context, mature integrations, and a security architecture built by the people who built the model. It describes the destination, not the starting line. For a mid-market company, that 65% isn't a projection of what you'll get next quarter. It's a picture of what becomes possible after the under-the-hood work is done — the work Karpathy flagged and most buyers are hoping to skip.

You don't buy a teammate. You build the conditions one can join.

Put those two things together — Karpathy's parenthetical and Anthropic's own architecture — and the strategic conclusion writes itself. The "AI teammate" is not a product you purchase and drop into your org. It's a capability that emerges when your organization has done the groundwork to let one function: your context is accessible, your tools are integrated, your data access is scoped, and your governance is deliberate.

This is the same truth we've been circling from every direction. It's why we've argued that your AI doesn't actually know your business until you've built the context layer that teaches it — and a taggable teammate with no memory of how your company works is just a very confident stranger in your channel. It's why the companies actually seeing returns from AI are the ones that made architectural decisions rather than tool purchases. Claude Tag doesn't change that thesis. It's the most vivid delivery surface for it yet: it takes the invisible work of context and integration and makes its absence painfully visible, because a teammate who doesn't know anything is obviously useless in a way a chatbot never quite was.

The honest framing for a leader is this: the exciting headline ("hire an AI teammate") is downstream of a quarter's worth of unglamorous work you may have been deferring — getting your context in order, wiring your integrations, deciding who and what the AI can see. The teammate is the reward for that work. It is not a substitute for it.

What "building the house" actually means

It's fair to ask what that groundwork actually is, because "build the architecture" can sound like a consultant's hand-wave. It isn't. There are four specific things that have to be true before an AI teammate is useful and safe in your company — and, tellingly, only one of them is about the AI.

The first is context — the teammate has to know how your business actually works. Not the public version; the real one: your products, your customers, your standards, the way a good proposal reads and a bad one fails. A teammate with no memory of any of that is a bright new hire on their first morning — every morning. Claude Tag builds some of this from a channel's history, but channel chatter is a thin substitute for the deliberate context layer that makes an AI genuinely useful. That's the whole difference between an AI that actually knows your business and one that's merely articulate.

The second is integrations — the teammate has to reach the tools where the work actually lives. A colleague who can talk but can't open your CRM, your project system, or your files isn't a colleague; it's a suggestion box. Every integration is real engineering, and every one is also a new door into your systems that has to be built and secured on purpose.

The third is scoped identity and access — the teammate should see only what its role should see, and no more. This is the part Claude Tag genuinely does well out of the box: a Claude configured for one channel can't carry its memory or reach into another. But the vendor gives you the mechanism; you still make the decisions. Which channels, which tools, which data — those are judgment calls about risk, and they're yours to make, not the product's.

The fourth is governance — a human owns the "yes." Someone is accountable for what the teammate does, and for anything it does on its own there is a point where a person reviews and approves before it acts. This is the layer most companies skip, and it is the one that turns "we deployed an AI teammate" from a story about productivity into a story about an incident.

Notice that three of those four have almost nothing to do with AI. Context, integrations, and access governance are the ordinary infrastructure of a well-run company — work that was valuable before Claude Tag existed and will be valuable long after. That is the reframe that should change how you budget. The line item that reads "AI teammate" is small and exciting; the work that makes it pay off is larger, older, and mostly about getting your own house in order. The companies that have done it will add a teammate and watch it thrive. The ones that haven't will add the same teammate, watch it flail, and — understandably but wrongly — conclude the AI was overhyped.

The mid-market reality check

I want to be specific about what this means for a 50-to-300-person company without a dedicated AI or IT function, because that's who we work with and the glossy version of this story does them a disservice.

Turning on Claude Tag is not a toggle. It requires a paid Slack plan, an Enterprise or Team subscription, and an owner to hand-configure which channels the AI joins and which tools it can use — deliberately, channel by channel. That's setup work, and it's exactly the kind of work that stalls in an organization where "the AI project" is somebody's third priority.

And there's a governance trap sitting right at the start. The guidance from people who've deployed it is blunt: do not add an AI teammate to channels carrying HR conversations, salary data, legal strategy, or client PII — anything you wouldn't want a new employee reading on their first day. That sounds obvious written down. It is not obvious at 4pm on a Friday when someone with admin rights is excited and adds Claude to the channels it would be "most useful" in. The failure mode isn't exotic. It's an ordinary access-control mistake, made by an under-resourced team moving fast — and it's the sort of thing that turns "we onboarded an AI teammate" into a data-exposure incident. This is the same discipline we've argued for around personal AI accounts and shadow AI: the answer isn't fear, it's deliberate access — decide what the AI sees on purpose, not by default.

There's a predictable shape to how this goes wrong, and it's worth picturing before you live it. The demo is flawless — someone tags Claude in a channel, asks for something, watches it work, and the room is sold. Then the rollout stalls, quietly, over the next month. The teammate keeps giving generically competent answers because no one wired in the context that would make them specific to your business. It can't finish half the tasks people hand it because the integrations to your real systems were never built. Someone hesitates to add it to the channel where it would actually help because they're not sure what it can see. And the enthusiasm curdles into "we tried the AI teammate; it was fine, nothing changed." None of that is a failure of the product — every one of those stalls traces back to a piece of groundwork that wasn't done. The demo tests the model. The deployment tests your architecture. Mid-market companies keep buying on the strength of the first and losing on the second.

The way through isn't more ambition; it's less. Pick one channel and one genuinely repetitive workflow — the weekly report, the first-pass research, the customer-reply draft — and do the unglamorous work to make the teammate excellent at that single thing: give it the context, wire the one or two integrations it needs, scope its access, put a human on the approving end. Prove it there, where "good" is easy to define and the data is safe, before you let it near anything broader. A teammate that's genuinely great at one workflow earns the trust and the budget to expand. A teammate dropped into ten channels to be mediocre in all of them earns neither.

The honest limits: architecture is necessary, not sufficient

Here's where I have to be straight, because the tidy version of this argument — "just build the architecture and the teammate works" — isn't fully true, and pretending it is would be exactly the kind of overclaim we tell clients to distrust.

Two real gaps remain even when the plumbing is excellent.

The first is a safety gap Anthropic hasn't closed yet. Claude Tag's ambient mode — where it acts proactively without being tagged — ships, as best we can tell from Anthropic's own description of the design, without a step where a human reviews and approves an action before it executes. An AI reading every message in a channel and acting on its own initiative is a genuine governance risk, and it's one that good architecture on your side doesn't automatically solve, because it's a property of the product. This is precisely why we keep insisting on keeping a human above the loop with real approval gates for anything autonomous. If you deploy an ambient AI teammate, the approval gate is something you have to build around it — it is not there by default.

The second gap is human, and no architecture reaches it. A developer named Daniel Nwaneri wrote the sharpest thing I've read about Claude Tag, and it has nothing to do with tools or memory. He pointed at the trust problem: when a colleague delegates real work to an AI, "good output gets read as 'she's outsourcing her thinking.' Mediocre output gets read as 'see, this is what we were worried about.'" That's an unwinnable social frame, and it's not a plumbing problem. You can scope every credential perfectly and still have a team quietly recalculating who's actually doing the work and whether it's allowed. This is the Humans First layer that the infrastructure conversation always misses: putting an AI teammate in the room changes how humans see each other's contributions, and that has to be led — named, normalized, made safe — not engineered.

It's worth staying with that trust problem, because it's the one leaders are least prepared for and the one most likely to quietly kill adoption. When you put a shared AI teammate in a channel, you don't just add a capability — you change the social arithmetic of the room. Every visible contribution now carries an implicit question: did a person do this, or did they tag Claude? Because that question has no clean answer, it corrodes in both directions. The careful employee who uses the teammate well looks, to a suspicious colleague, like someone coasting. The one who refuses to touch it looks, to an impatient manager, like someone falling behind. Neither read is fair, and both go live in your organization the moment the teammate joins the channel.

No amount of per-channel scoping touches that. It's a leadership problem, and it has to be led out loud. Three moves matter. First, make using the teammate the expectation, not the confession — say plainly that tagging Claude is how work gets done here, so it stops being something people hide and becomes something they compare notes on. Second, decouple "who did the work" from "who's accountable for it": the person who delegates owns the output completely, exactly as a manager owns work they assigned to a junior. That reframe dissolves most of the "is she outsourcing her thinking" anxiety, because it moves the standard from effort to result. Third — the one leaders skip — model it. If the executive team quietly avoids the teammate while telling everyone else to embrace it, the organization reads the avoidance, not the memo. The fastest way to make an AI teammate normal is for the most senior person in the channel to visibly delegate to it and visibly own the result.

There's a deeper thing underneath the trust problem, and it's where the value in all of this is actually moving. When a teammate can do the doing, the human contribution shifts from producing the work to judging it — deciding what to delegate, catching what the AI got confidently wrong, owning the call. That's a real contribution, arguably the more valuable one, but it's less visible than typing, and organizations have spent a century learning to reward visible effort. A team that still equates contribution with visible output will systematically undervalue its best people the moment they start delegating well — and quietly reward the ones performing busyness. So the trust layer isn't only about permission to use the tool. It's about updating what your culture counts as contribution in the first place. That's not an IT rollout; it's a leadership act — and it's the same shift underneath every one of these pieces: the scarce human input is judgment, and a company has to learn to see judgment as work.

And notice that the trust question and the money question are the same question in two hats. A shared teammate can run up cost on someone's tag while no one feels like they're "using" it, and it can produce work no one is quite sure they can claim. Both are failures of ownership, and both are fixed the same way: by making explicit, per piece of work, who owns the output and who owns the spend. Assign the teammate the way you'd assign a shared resource with a budget and a supervisor — not the way you'd install an app. The companies that name ownership early will look back on it as the boring governance decision that made everything else possible; the ones that don't will meet both problems in the same month — one on the invoice, one in a tense team meeting.

So the honest thesis is not "architecture solves the AI teammate." It's this: architecture is the necessary condition, and it's the part most companies skip — but it isn't the whole job. The plumbing makes a teammate possible and safe to have in the room. Leadership makes it something a team actually trusts and uses well. Skip the first and you have an expensive stranger; skip the second and you have a perfectly-scoped tool nobody trusts.

The quieter shift underneath: what you're actually buying

There's one more thing leaders should see coming, because it changes the economics. The investor Saanya Ojha noted that a shared, taggable teammate breaks the per-seat mental model we've all been using for software: "a shared agent can be invoked by many people… and generate cost while no one is actively 'using' it." You're no longer buying seats. You're buying throughput — work the teammate does, sometimes on its own initiative, sometimes overnight, whether or not a human is watching.

That's a governance and ownership question dressed as a pricing change. Who owns the AI teammate's output? Who's accountable when it acts in ambient mode? What's the budget for work that happens without a person initiating it? These aren't reasons not to adopt — they're the questions that separate a deliberate rollout from a surprise on next quarter's bill. The teammate framing is genuinely apt, and it cuts both ways: you onboard a teammate with scoped access, a probation period, and oversight — you don't hand a new hire the keys to every room on day one and let them act unsupervised. The same is true here.

What to do before you "hire" the AI teammate

If you're tempted by Claude Tag — and you should be; the direction is real — treat it like onboarding a teammate into a well-run house, not installing an app. Four moves, in order:

Do the under-the-hood work first. Before the teammate joins, get the conditions right: the context it needs to actually know your business, the integrations to your real tools, and a clear map of what data lives where. This is the quarter of unglamorous work Karpathy flagged. It's also the work that pays off across every AI initiative, not just this one — so it's not a Claude Tag cost, it's an AI-readiness investment.

Decide access on purpose, channel by channel. Start narrow. Put the teammate in the channels where the work is genuinely useful and the data is genuinely safe for a new participant to see. Keep it out of HR, legal, finance, and PII channels until you've deliberately decided otherwise. Default-closed, opened on intent.

Put a human at the gate for anything autonomous. If you use ambient/proactive mode, build the approval step the product doesn't give you — a human who reviews and owns the action before it lands. Autonomy without an approval gate is not a feature; it's an unmanaged risk.

Lead the trust question openly. Name it with your team before it festers: using the AI teammate is encouraged, not a sign someone's coasting; the human is accountable for the output regardless of who drafted it. The social frame Nwaneri described only becomes toxic when it's left unspoken. Say it out loud and it loses its charge.

The teammate is real. So is the work.

Karpathy is right that this is a genuine shift — the third interface, the move from tool you operate to teammate you delegate to. It's not hype, and the companies that get there will have a real edge. But the same person who named the paradigm named its price, and it's worth ending on his word rather than mine: "once you do all of the under-the-hood engineering work."

That clause is the whole game. The teammate is the part you can see and the part that's easy to want. The engineering underneath — the context, the integrations, the scoped access, the governance, and the leadership around all of it — is the part that decides whether the teammate is a colleague or a liability. You don't buy your way to an AI teammate. You build the house first, and then it can move in. The companies that understand that will spend this year doing the quiet work. The ones that don't will add a bot to a channel, watch nothing change, and conclude the paradigm was oversold — when what was actually missing was the floor they never built.

Frequently Asked Questions

What is Claude Tag?

Claude Tag is Anthropic's feature (launched June 23, 2026) that lets Claude join a Slack channel as a shared, taggable team member. Admins grant it access to specific channels, tools, and data; anyone can @-mention it to delegate a task, and it works through the task in the open thread. It builds memory from channel history and, in "ambient mode," can act proactively. It's currently in beta for Claude Enterprise and Team customers, on Slack first.

What did Karpathy mean by the "third paradigm" of AI?

Andrej Karpathy described Claude Tag as the "3rd major redesign of LLM UI/UX": the first paradigm was the LLM as a website you visit, the second was an app you download, and the third is a persistent, asynchronous entity with org-wide tools and context that works alongside your team. Crucially, he also named the precondition — the "under-the-hood engineering" (tools, integrations, compute, memory, security) required to make it actually work. It's his framing, not Anthropic's official one.

Is an AI teammate something you buy or something you build?

Both — but the hard part is what you build. You can subscribe to Claude Tag, but it only becomes a useful teammate once your organization has done the underlying work: accessible context, integrated tools, scoped data access, and deliberate governance. The interface is purchased; the conditions that make a teammate effective and safe are built. That's why Anthropic's own "65% of code" stat describes a mature environment, not a starting point.

Is Claude Tag safe to use with company data?

It's designed with real safeguards — per-channel identity scoping, sandboxed credentials injected at request time, default-deny networking — but safety still depends on how you deploy it. The key risk is access: don't add an AI teammate to channels with HR, salary, legal, or client-PII data ("anything you wouldn't want a new employee reading on day one"). And note that ambient/proactive mode ships without a built-in human approval step, so you should add one for anything autonomous.

Does Claude Tag work for a mid-market company, or just big tech teams?

It can, but it takes deliberate setup: a paid Slack plan, an Enterprise/Team subscription, and hand-configured per-channel access by an owner. For a company without a dedicated AI/IT function, that's real work — and the payoff depends on having your context and integrations in order first. Anthropic's internal results don't automatically generalize; they reflect an environment already built for this.

What's the risk of an "AI teammate" beyond security?

The human trust layer. When someone delegates work to an AI, colleagues may read good output as "they're outsourcing their thinking" and weak output as confirmation the whole idea is bad — an unwinnable social frame that no architecture fixes. It has to be led: make clear that using the teammate is encouraged and that the human remains accountable for the output. Left unspoken, the trust problem quietly undermines adoption.

How is this different from a regular chatbot or the old Slack AI app?

A chatbot is a private, one-to-one tool you operate turn by turn. Claude Tag is a shared, persistent participant in a channel — visible to the whole team, working in the open thread, remembering context across time, and able to act proactively. That shift, from a tool you use to an entity you delegate to, is what Karpathy means by a new interface paradigm — and it's why the governance and architecture questions are bigger than they were for a chatbot.

Sources

• Anthropic — Introducing Claude Tag (June 23, 2026) — official launch; Slack teammate, per-channel access, ambient mode; "65% of our product team's code" internal stat
• Andrej Karpathy — X post (June 23, 2026) — the "3rd major redesign of LLM UIUX" framing and the under-the-hood engineering precondition
• TechCrunch — Anthropic's Claude Tag is learning your company, one Slack message at a time (June 23, 2026) — per-channel identity scoping; product detail
• VentureBeat — Anthropic launches Claude Tag, a persistent AI teammate (June 23, 2026)
• Fortune — Anthropic releases Claude Tag, a virtual employee that works within Slack (June 23, 2026)
• Bloomberg — Anthropic Wants Claude to Be Your New Slack Coworker (June 23, 2026)
• Security Boulevard — Claude Tag Security: Governing AI Identities in Slack (June 2026) — sandboxed credentials, default-deny networking (reported)
• Daniel Nwaneri — Everyone's Excited About Claude Tag. Nobody's Built the Trust Layer. (DEV Community, June 24, 2026) — the trust-layer critique
• Saanya Ojha — Tag, You're It (The Change Constant, June 24, 2026) — the per-seat → throughput economics shift