When the Agent Joins the Channel
Did you see last week that Anthropic released a Slack agent they call Claude Tag? This means that Claude is no longer only a private chatbot in a separate window, waiting for someone to leave the conversation, open a tool, and ask for help. With Claude Tag, it can sit inside Slack itself, functionally as a member of the channel. A person can tag it in a thread, and Claude can answer where everyone else is already working. It can follow the conversation as it moves, use the tools the team has connected, remember channel and workspace context over time, and respond without a fresh mention or run standing routines that surface updates back into the channel. The agent is no longer off to the side, waiting to be consulted. It is in the room where the work is being made.
A year ago, writing about banks that had started giving AI agents email addresses and Teams logins, I asked what would happen when one of those agents joined the firm’s Slack channel and became a working member of the team. Claude Tag is that question answered.
The difference is position. The agent is not a window you open and close when you need something. It is part of the same running conversation as everyone else. Someone drafts a paragraph and asks the agent to check the cites. Someone else disagrees with the framing. The agent is asked to hold the open issues, recall what the thread decided three days ago, or turn a long argument into a short list of what still needs to be resolved. It is not receiving the work after the team is finished thinking. It is watching the thinking happen.
It will feel ordinary almost at once. A lawyer who would never open a standalone AI product may still tag the agent because everyone else in the thread already is. A staff member who does not think of herself as an AI user may rely on the channel summary because that is where the answer now lives. Nobody has to announce a grand adoption strategy for this to change behavior. People will just keep up with the conversation, and the conversation will now include the agent.
After a few weeks of working this way, the team will stop narrating it. People will tag the thing by whatever name they have given it, Bob or otherwise, and wait on it like a colleague who is useful, dependable, and occasionally a little slow. The morning you catch yourself irritated that Bob is taking too long to answer is the morning you have accepted that Bob is on the team.
None of that means the agent is in charge. It means the work begins to flow through a new participant. Its draft goes into the file. Its summary becomes what people refer back to. Its list of open issues becomes the list everyone works from. And it does all of this under its own name. In a channel the agent acts through its own accounts and leaves its own audit trail, so the work is attributed to the agent, not to whoever tagged it. That creates real questions. Who owns the language? Who checked the record? Who verified the citations? Who made the judgment call before the work left the channel? Those questions are no longer hypothetical. They are the cost of letting a new teammate carry real work, and they come due whether or not anyone stops to name them.
There is a deeper change underneath this one, and it is the one I want to take up next week. An agent that sits in the room long enough does not only help with the work. It starts to learn how the room works, and that may matter more than anything it drafts.
For now, the point is smaller and stranger. We spent the first wave of legal AI deciding, person by person, whether to use the tool. That decision is beginning to leave the table. Once the agent is added to the place where the work already happens, the room does the rest.
The agent did not just get better. It got a seat.
