THE SIGNAL
The Coworker That Gets Harder To Fire
Anthropic is not shipping a better Slack bot. It is installing an agent whose value grows with everything it overhears, which is the same thing that makes it expensive to remove.
Anthropic Retires The Bot And Hires An Agent
What happened: Anthropic is shutting down its existing Claude in Slack connector (the simple app that let workers ask the Claude AI questions inside Slack) on August 3, and replacing it with an always-on agent called Claude Tag. Tag joins a company's Slack as a team member, reads whatever channels administrators give it access to, and builds up its own understanding of the business from everything that flows through those channels. Enterprise and Teams customers get access now. To pull companies across, Anthropic is handing out migration credits worth $25,000 to Enterprise accounts and $2,500 to Teams accounts that move before the credits expire on September 1.
What's really going on: The headline feature is persistence, and persistence is the business model. Tag learns the longer it sits in a channel, pulls in context from other channels and data sources it is allowed to touch, and can schedule its own work to run over hours or days. Every week it operates, it accumulates a private picture of how your company actually works, and that picture lives on Anthropic's servers. That accumulated context is the product, and it is also the exit cost. Leaving is not swapping one bot for another. It is throwing away months of learned institutional memory and starting any competitor from zero. Pair that with Anthropic's recent move to metered, pay-per-token pricing, and an agent that pursues its own assignments around the clock stops being a convenience. It becomes a meter that never stops running.
Why most people are missing this: They are reading this as "Claude in Slack got smarter," when the real change is a switching cost that grows on its own the longer you leave it alone.
The Take: The credits are not a discount. They are the cost of getting the agent embedded before anyone does the math on what it would take to pull it out.
Why it matters: Once an agent has learned a company's channels, its code, and its decisions, the question stops being "is this the best model" and becomes "can we afford to retrain a replacement on everything this one already knows." That is a moat made of your own data, and you are the one filling it.
The Pattern
The tension is between AI as a tool you call and AI as a coworker you embed. A tool is stateless, cheap to drop, and easy to compare against rivals. An embedded agent that learns your context is none of those things, and that is exactly why vendors want to ship the coworker, not the tool. The side winning is the one that turns daily usage into accumulated, non-portable context, because in a market where the underlying models are at most a few months apart, that context is the only durable lock-in left.
What This Signals
The fight is moving from model quality to context custody. Whoever holds the learned history of how your company runs holds the account, no matter who ships the better model next quarter.
Metered pricing plus a self-scheduling agent turns "always-on" into "always-billing," and the customer has quietly handed the judgment of how much work needs doing to the thing that profits from doing more of it.
A migration deadline and expiring credits look like a promotion. They are really a mechanism to embed the dependency during the short window before buyers price the exit.
Quick Byte
King Camp Gillette is remembered for giving away razors to sell blades, but he never actually did it. The cheap-razor, expensive-blade model only took hold after his patents expired in 1921, when rivals proved that recurring consumption beat the one-time sale. The free hardware was always about the refills.
THREAD
Anthropic is retiring its Claude in Slack bot for an always-on agent that learns from every channel it sits in. The learning is the feature. It is also the reason you cannot easily leave.
An embedded agent that accumulates months of your company's context cannot hand that knowledge back. Switching vendors means throwing the memory away and training a rival from zero. That is the moat, and you built it.
If the agent schedules its own work around the clock and you pay by the token, who actually decides how much work needs doing?
POST: Anthropic is replacing its Claude in Slack bot with an always-on agent that learns from everything in the channels it joins. That learned context sits on Anthropic's servers and does not come with you if you leave. The longer it runs, the more it knows, and the more it costs to walk away. The $25,000 in migration credits is not generosity. It is the price of getting embedded before customers calculate the exit. The model is months from being matched. The lock-in is not.
TAKE: The strongest AI lock-in is not a better model. It is an agent that has quietly learned your entire company and has no way to hand that knowledge back when you go.