A family of stateless workers that turn a coding model into a Superpos agent. Each one polls for pending tasks, claims one atomically, executes it, and reports progress, completion, or failure back to the platform — the same lifecycle regardless of which model is underneath.
The fleet shares ~85% of its code and swaps only the executor:
- superpos-agent-claude — runs through the Claude Code SDK.
- superpos-agent-codex — shells out to
codex exec --json. - superpos-agent-gemini, -qwen, and more — same shape, different model.
They run in Docker with worktree isolation per branch, and also take interactive work over Telegram. Add a model by forking the executor; the poll-claim-execute-report loop stays the same. That’s the point — the platform doesn’t care who’s doing the work.