The agent in the room has to introduce itself.

On Tuesday 10 November Anthropic announced Claude Meetings. Bring it into a video call and it joins as a participant you can see, listens, and puts the notes, action items and anything it worked on during the call into claude.ai once the call ends. It launched as a beta for Team and Enterprise plans, with Google Meet the only supported platform so far, alongside new whiteboard and spreadsheet tools.

Meeting notes from software are not new. What caught our eye was the small print. According to the launch coverage, the agent posts a notice in the meeting chat saying what it is. It will not join a call with external guests unless an admin has allowed it. And if an outside guest joins partway through, it leaves on its own.

Those three rules are the actual product. They are also the rules we keep arriving at whenever software is put into a space that people share.

Presence is a design decision

A tool you open is yours. An agent that sits in a room is everyone's problem. The moment software is present in a shared space, whether that is a call, a reception desk, a shop floor or a brand installation at an event, the people around it did not all choose it. Some of them do not know it is there.

The cheapest way to lose trust in a system like that is for someone to find out afterwards. So the first job is a plain, visible statement of what is present and what it is doing. Not a line in a privacy policy. Something in the room itself, at the moment it matters.

"

If people only find out the system was listening after the fact, the system was designed wrong, however good the notes are.

The room changes, so the rules have to

The rule we like most is the one about leaving. An internal call and a call with a client are different rooms, even if the link is the same. Software that keeps behaving the same way when the audience changes is the source of most of the uncomfortable moments in this area.

We see the same pattern outside meetings. An interactive installation that captures photos needs to behave differently when it is clear a child is in frame. A kiosk that remembers a returning customer needs to stop remembering when someone else steps up. A staff tool that is fine in the back office should not be readable from the counter. In every case the system needs to know who is there, and step back when it cannot be sure.

Who said yes

The admin approval rule answers a question every agent eventually raises: who agreed to this? One person bringing an agent into a call should not be able to decide on behalf of a client in another company. Consent belongs to whoever owns the space, and someone with the authority to grant it has to do so deliberately.

For the systems we build that usually means two layers. The organisation decides where an agent or a capture feature is allowed to exist at all. The people in the moment can see it and, where it matters, turn it off. Neither layer on its own is enough.

What we would do

• 01Make presence visible where the people are, not in a settings page. A notice in the chat, a light on the device, a line on the screen.
• 02Decide in advance which rooms the agent is allowed into, and make that an admin decision, not a per-user one.
• 03Treat a change of audience as a trigger. When someone new arrives who did not agree, pause, leave or stop capturing.
• 04Keep what was captured somewhere people can see and delete it. An agent that writes things down is now a records system, and should be built like one.
• 05Pilot it internally first. Rooms full of colleagues are forgiving. Rooms with clients in them are not.

Agents are going to turn up in more of the places we design for, not just calls. The ones people accept will be the ones that introduce themselves, ask the right person before they stay, and know when to leave.