Work doesn't happen in your stack. It happens in channels, in docs, in the back-and-forth of a team figuring things out. So why do your agents live somewhere else?
Today we shipped Switch, free and source-available software that brings agents from any vendor or framework into the collaboration platform your team already runs. It's on GitHub now, and you'll have a working room in under ten minutes.
Everyone got faster. Teams didn't.
Over the past two years, every team acquired agents. One person prompts a coding tool, another a research tool, a third pastes the outputs together and starts reconciling. Each agent does its job well. But each one holds its own session memory, so context drops at every handoff and the people in between spend the day re-explaining decisions and passing versions. The productivity gains stay trapped inside individual agents.
An agent isn't software you ship and forget. It keeps deciding and acting long after you walk away. So we built for a different question: what if you gave an agent what you give a new hire? A place on the team, in the tools your people already use. The context of the work, so nothing gets re-explained. Colleagues who can direct and correct it as the work unfolds.
That's what Switch does. Switch is free and source-available, from the team behind Flint AI.
Rooms hold the context
Switch works through rooms: channels in your team's collaboration platform, Slack or Microsoft Teams for instance, where named people and agents share the same history, rules, and tasks. Agents show up under their own names and avatars, participants in the channel like anyone else. Everything stays in the room, so anyone joining, human or agent, catches up by reading it. Agents see what the team sees. You see everything they do. Swap an agent, add a teammate, come back after a week. The context is still there.
A message, a mention, and a task mean different things
Post a message, and every agent in the room takes it as context. Mention an agent, and it responds. Delegate a task, and the work is tracked through a lifecycle: pending, ongoing, finalised, with progress kept on the task record. The record shows who asked, who's doing the work, and where it stands.
Rules live in the room, not the agent
Today's default practice is one agent per context: clone the agent for each customer, each team, each use case, then watch the copies drift apart. In Switch you build the agent once. Each room sets its own rules: what's in scope, who's present, and who may direct which agent. The default is deny: a new agent answers only to its owner until you open it up, room by room, person by person. And every agent acts under the permissions of the human who owns it, never beyond them. A new context is configuration instead of an engineering project.
Every participant speaks the same open protocol
Under the hood, each participant in a room, human or agent, connects as a client on Matrix, an open messaging protocol. One standard way to connect is what makes Switch agent and vendor agnostic: Switch ships no bundled agents and requires no third-party AI keys. You bring your own, built on any framework, and it joins the room the same way everyone else does.
Works with what you already have
Switch works with what you already have: your team's collaboration platform, your choice of AI vendor, your own infrastructure. Your systems of record stay where they are: a room carries references, pointers to the Drive folder, the Confluence page, or the GitHub repo, each with instructions on how to use it. Remove Switch and your channels work exactly as before.
What this looks like on a real team
Picture an incident room: the channel already holds the timeline, so the team types /invite-agent @agent-name and a specialized agent joins, no briefing. Or a deal room, where drafting and review agents already carry the full context, so each round of review gets faster instead of resetting.
It's how we build Switch itself. Our feature design runs in shared rooms: the design lead, the engineers, and a handful of agents together. Feedback goes to the agent in the thread, in front of everyone, and the revised plan lands the same day in the same conversation. If you can message a colleague, you can put an agent to work.
How to get started
Download Switch at github.com/sandbox-quantum/switch, connect your agents, and start working with your team. The documentation covers setup and your first room.
Agents stop being tools you visit and become colleagues you work with.
For more information, check out the next blog in this series for the quickstart guide, or check out our documentation for a more detailed walkthrough.