Switch：在团队聊天中协作使用AI代理

Ask the agent to find the bug. Read the answer. Copy the useful part. Paste it into Slack. Explain what you already tried. Answer your teammate's questions. Go back and tell the agent what your teammate said. I have run that loop more times than I can count. Every pass lost a little detail, and the agent session ended up as the only complete record of the investigation. I spent two evenings this week testing Switch, an open source, self hostable project that puts the agents you already use into shared rooms on Slack, Teams, Discord, Mattermost and Telegram. People and agents work in the same thread, so the history and instructions belong to the room, and anyone who joins can read them. 𝗪𝗵𝗮𝘁 𝗜 𝘀𝗲𝘁 𝘂𝗽 - A room for our content scheduling service, bridged to the #scheduler channel in Slack - Claude Code connected through Switch Console - Room Instructions: what the room is for, who owns what, how agents report progress - A reference to the repo so agents know where the code lives The detail that matters to me, since I work on agent oversight: a reference tells an agent where something lives, and the agent can still open only what its own credentials reach. Attaching a repo to a room grants nobody new access. 𝗧𝗿𝘆 𝘁𝗵𝗶𝘀: 𝘁𝗵𝗲 𝗳𝗿𝗲𝘀𝗵 𝗮𝗴𝗲𝗻𝘁 𝘁𝗲𝘀𝘁 The Switch docs describe a check I now run on every room. 1. Write the room Instructions in Switch Console. 2. Invite an agent that has never been in the room. 3. Give it no background. Ask what the room is for. 4. Read the answer. A pass uses your team's names for things and describes the actual job. An answer that would fit any room means the Instructions are empty or the real guidance is buried in a document. My first room failed. The new agent gave me a generic summary about software projects, because I had put the real guidance in a room document and left the Instructions almost empty. I moved the ownership and naming rules into the Instructions, invited a second fresh agent, and it described the room correctly on the first reply. 𝗪𝗵𝗮𝘁 𝗰𝗵𝗮𝗻𝗴𝗲𝗱 𝗳𝗼𝗿 𝗺𝘆 𝘁𝗲𝗮𝗺 On day two a teammate joined the thread cold, scrolled the history, and asked the agent why the retry job was skipping scheduled posts. The agent answered from the investigation already in the room, and I did not have to write a summary for anyone. The repo has examples for Claude Code, Codex and OpenCode, plus a documented protocol for custom agents, so the team can change agents later and keep the room. Repo and setup guide - http://devtoolsacademy.link/brij If you build a room, show your workflow with #HowISwitch. When an agent joins a team, what should its onboarding include? Thanks to Switch for sponsoring this post.