humans& Launches Boba, a Group Chat AI Trained on Social Restraint

AI lab humans& has launched Boba, a messaging assistant designed to coordinate people around shared plans. Boba runs on Ampersand, a model trained for multi-person collaboration that can unfold over weeks or months.

You can text Boba one-on-one or add it to a group chat. It can create reminders and calendar invitations, send email, research the web, and help participants track commitments. Early users have exchanged more than one million messages with the assistant, according to the company.

Quiet is part of the training

Conventional model training often rewards completion of bounded tasks and favorable ratings for individual responses. Humans& argues that those objectives can produce assistants that intervene too often, overlook group dynamics, and optimize the immediate request without accounting for its social consequences.

The lab calls its approach Reinforcement Learning from Simulated Feedback, or RLSF. Ampersand trains inside simulated communities, interacting with thousands of modeled people across millions of simulated days. That extended environment allows the reward signal to incorporate the consequences of an action long after the model takes it.

A separate user model called Persimmon generates the simulated participants. It varies their preferences, attention, and behavior, accounting for people who become distracted, resist suggestions, or change their minds.

Evaluation follows a similar process. The team places Ampersand in simulated groups with goals grounded in real data, allows those goals to evolve over as much as several simulated months, and measures progress toward them. The score subtracts costs for interactions that conflict with user preferences.

The team reports that this training produces social restraint. In a group chat, Boba is designed to remain silent during casual banter and contribute when it has useful information, a relevant action, or a coordination task to handle.

Early use centers on practical chores

Reported uses focus on recurring tasks where missed details and weak follow-through create problems:

• Subscription review: Boba read the terms of a supplement trial, identified an automatic conversion to a paid subscription that the user had missed, and helped cancel it.
• Job searches: Users worked with Boba to revise resumes, access job boards, and submit applications.
• Personal administration: The assistant handled meeting reminders, daily briefs, follow-ups, medication check-ins, doctor appointments, and diet records.
• Group coordination: Chats used Boba to plan trips, compare prices, organize family logistics, and create images or songs together.

An internal survey drew 290 responses, with 70% saying they loved Boba and 25% saying they liked it. The survey provides an early satisfaction measure from a small company-run sample; broader adoption and long-term retention will require larger cohorts and behavioral data.

Boba exposes the underlying stack

Two days before launching Boba, the lab released Ampersand v0.1 through an Agents API for developers building collaborative applications. Boba serves as the company’s consumer implementation of that stack.

The developer documentation describes a session-based runtime with a straightforward division of responsibilities:

• Conversation state: Ampersand runs the agent and retains its conversation history.
• Messages and tools: The application sends messages and supplies the tools the agent may call.
• Reply routing: The application decides where and when the model’s responses appear.

This architecture gives developers a collaboration-focused alternative to adding social instructions around a general-purpose language model. Production integrations will still require explicit choices about tool authorization, data retention, group consent, and failure handling, especially when an agent can send email or create calendar events for multiple participants.

OpenAI, Anthropic, and Google have made autonomous agents a major product direction. Humans& is pursuing coordination among people, supported by long-horizon memory, multi-agent systems, and user modeling. Researcher Andi Peng described those capabilities as central to the lab’s training agenda when the company launched.

Simulation has to survive reality

For researchers, RLSF offers a concrete test of whether simulated multi-agent environments can teach durable social behavior. Standard reinforcement learning from human feedback commonly evaluates individual responses; RLSF extends the scoring window across evolving goals, relationships, and repeated interactions.

The approach depends on transfer from modeled communities to real groups whose preferences and relationships may exceed Persimmon’s simulations. Independent evaluation would help establish whether Ampersand’s reported restraint and coordination gains persist outside the lab’s environments.

Boba is available in several countries, including the United States, Canada, and Australia. Login and browser access remain alpha features. Its practical benchmark will be continued use: whether groups keep Boba in their threads after the novelty fades and trust it with enough access to make its coordination tools useful.