A team built 'Shelly', a remote AI coding agent system where a Slack message triggers a fully provisioned, ephemeral Fly.io machine running Claude Code via an Elixir/Erlang cluster. Each agent gets its own isolated machine, uses the team's devcontainer environment, and communicates back through an Erlang cluster where a primary 'Admiral' node handles privileged operations like Slack tokens — keeping sensitive credentials off agent machines. The system uses FLAME for machine lifecycle management and Erlang distribution as the coordination layer. After 5 weeks, 40% of PRs originate from Shelly, with 7 non-engineers having PRs merged, averaging 17 agent-originated PRs per week.
Table of contents
Remote Agents are the next stepEach Task Gets Its Own MachineThe Erlang Cluster Is the SuperpowerWhere we are todaySort: