How I Put Three AI Agents in a Discord Channel with Switch (and What Broke on Windows)

ShipLoop, a project developed for the Flip the Switch hackathon, links customer complaints in Discord to GitHub pull requests, with a human making the final merge decision. This account explains how to set up the project and the challenges faced when running Switch on Windows.

The process involves three AI agents—Triager, Fixer, and Reviewer—working together in a Discord channel. Triager classifies the issue and hands it off to Fixer, which reproduces the bug, writes a failing test, makes the fix, and opens a pull request. Reviewer checks the pull request with its own edge-case tests. Once merged on GitHub, Triager reports the elapsed time and drafts a customer update for approval.

Switch is a tool that places AI agents in team chat, allowing them to function like teammates without the tool shipping its own agent or model. To set up ShipLoop, install the necessary tools, including Switch Console, Docker Desktop, and an AI coding agent like Claude Code. Start Docker, then use Switch Console to add a server and run it on the computer.

Test one agent using the Switch expert template and create additional agents by specifying their names, working directories, and instructions. Connect Discord by creating a bot in the Discord Developer Portal, granting it the required permissions, and inviting it to the server. Add room rules to control the workflow, such as limiting merges to humans and requiring approval before sending updates to customers.

Issues encountered during setup included the Claude Code CLI not being installed and port conflicts within Windows. Fixing these issues required reinstalling Switch Console and managing Windows port reservations by stopping and restarting the Windows NAT service. Additionally, connecting agents to multiple rooms caused communication problems, which were resolved by moving all agents into a single room. Another issue was the bot lacking the Manage Webhooks permission, which was addressed by granting it accordingly.

In summary, ShipLoop successfully merged two pull requests using the three AI agents, with a total run time of 6 minutes and 59 seconds. The project demonstrated that starting with one agent and one channel is advisable, followed by gradually adding more agents and rooms.