A group assistant needs to decide whether a contribution is useful before it decides what to say. Ampersand v0.1, from Humans&, makes that participation decision central to its design. Vercel announced its availability on AI Gateway on November 6, 2026, with free access through November 13.
A model built to join a conversation
Vercel says Ampersand was built on GLM 5.3 and trained in simulated group conversations. It reads messages and can choose to contribute or wait. The integration exposes caller-defined actions named send_message and pass, which the application executes and records. Vercel’s launch post
The gateway’s model page lists the API identifier humansand/ampersand-v0.1 and integration examples. Developers should consult the live provider listing for current availability and pricing as the introductory period ends. Model listing
Good timing is an evaluation target
A useful pilot would replay an approved set of team conversations and ask reviewers to score participation separately from answer quality. Did the assistant notice a contradiction? Did it interrupt a discussion that was progressing naturally? Did it wait for an answer after asking a question? These are observable behaviors that a generic correctness benchmark can miss.
Consider a launch-planning discussion in which one colleague mentions being away on Friday and another later schedules a Friday review. A helpful intervention connects the two messages and asks for a revised time. Repeating everyone’s messages or proposing a complete new plan would create extra reading without resolving the immediate conflict.
The same assistant can encounter private goals, informal remarks and decisions that only one person is authorized to make. Applications should define who belongs to a room, what history the model receives and which actions are allowed there. Membership changes should update that access before the next contribution. The conversation itself should not become an implicit grant to send messages somewhere else.
Make the application own delivery
The application-side actions are an important integration boundary. A model can propose a contribution while the host decides where to deliver it and records the result. A production implementation should use a stable request identifier so a retried action does not produce the same message twice.
Reviewers also need to see why silence was useful. Track occasions where the assistant correctly waited, as well as missed opportunities. Counting only sent messages can reward an overly talkative system and obscure the behavior the model is intended to improve.
For teams experimenting during the introductory period, a small internal room with consenting participants is a sensible first setting. Define the assistant’s role, evaluate a mix of routine and ambiguous conversations, and compare interruption rates with useful interventions. The result should tell the team whether group participation actually reduces coordination work.