Build, fix and review code with a chat-based crew of six role-bound agents

Justin and Ifeoluwa (Team Dynamo) built DevCrew for a hackathon: six agents on OpenCode that sit in a Mattermost team chat run locally by a tool called Switch, all defined in one Switch template. Each has one job. Scout reads the repo and runs tests but cannot edit. Conductor routes requests. Forge writes fixes on a branch. Sentinel re-runs the tests on its own and tries to break the change. Pulse writes standups from git and chat history, and Kudos credits contributions. An @mention in a message starts that agent, so Forge ending its report with a mention of Sentinel kicks off the review with no human typing. A person decides what ships. They pointed the crew at a small task-list app with two bugs they had planted.

Scout found both bugs, Forge fixed them and Sentinel confirmed the fix: the suite went from 21 passing and 2 failing to 25 passing and 0 failing. Plenty went wrong on the way. Forge answered questions outside its role until every agent got explicit boundaries, and it left its work uncommitted because nobody told it to commit. Stray @mentions inside task text woke agents for nothing. Agents bounced users between each other until every redirect pointed to Conductor. Forge could not open a pull request because the GitHub CLI was not signed in, so it handed over a link instead.

In this setup any mention starts an agent, so only real handoffs should contain one. Sentinel reported concerns more honestly with three verdict levels instead of two. The human approval step lived only in a prompt, not in the tool. This was a hackathon demo on planted bugs, not production use. Code and agent instructions are public in the linked repo.