A messaging app for you and your AI agents

A Billion Agents is where one person works with many AI agents. Each agent has a job, such as paying bills or researching a question, and you talk to it in conversations, the way you message people. It keeps notes and files of its own, so it remembers what you told it, and it works only when you message it.

It is in private beta, by invitation, and pricing has not been decided. Join the waitlist and we will email you when there is room.

The app with the conversation list on the left and a conversation open on the right, in which Chief of Staff answers questions about planning a team offsite.

What it does today

Conversations with your agents

Every conversation is in one list, newest first. A conversation can have one agent or several: the first one leads and answers, and tagging another by its handle, such as @researcher, brings it in to answer once. Starting a conversation always starts a new one, and you can pick up any earlier one again. Replies stream in as they are written, and Stop ends one you no longer need.

Agents that keep what they learn

Each agent has a workspace of text files: its instructions, its skills and its knowledge. When you tell an agent something worth keeping, it files a note in its knowledge, and it reads its files when they bear on what you ask. An agent works in one session at a time, and every conversation it is in feeds that session, so what you say in one conversation it knows in the next. Agent info shows what an agent is for, what wakes it and the files it keeps.

Agent info for Bill Payer: the time it worked this month, that it wakes up when you message it, and its knowledge files, such as knowledge/payables/recurring.md.

A new agent from a description

To make an agent, describe its job in a message. It drafts itself as a card with a name, a purpose and the files its workspace starts with; you change it by replying, and Create starts it. Ask an agent to set up another one for a project and it creates a delegate, which works on its own and reports back to the conversation you asked in.

A new agent's conversation: a request to keep track of invoices, a draft card for Invoice Payer, and after the reply asking to call it Invoice Chaser, a second card with a button to create it.

A computer, a browser and GitHub

An agent can start a computer of its own, run commands on it and copy files between the computer and its workspace. The computer pauses when it sits idle and is discarded when the agent's session ends. An agent can also read web pages in a real browser and carry out a task on a site; when the site asks for a sign-in, it hands the browser to you to sign in yourself. Before a step that is hard to undo, such as placing an order, it asks you, and acts only if you allow it. Connect a GitHub account, and agents can work with its repositories, issues and pull requests through it.

See how an agent works

Agents run on Claude, Anthropic's model. The inspector shows the instructions an agent works from, the files in its workspace, each run with the time it took, and what the agent is doing now. You do not need it to use the app; it is there when you want to check an agent's work.

The inspector for Bill Payer on its Runs tab: each run with what started it, what it did, when, and how long it took.

The private beta

A Billion Agents is in private beta: people join by invitation, from the waitlist. Pricing has not been decided. The app runs in a web browser, on a laptop or a phone. Agents work when you message them, or when one agent hands another work; they cannot yet act on a schedule or receive email.