← ClaudeAtlas

parallel-agentslisted

Orchestrate multiple AI coding agents in parallel terminal panes — fan work out (e.g. "fix this bug 3 ways", "have Claude/Codex/Gemini each take a stab"), watch their progress, and reconcile results. Uses Unterm MCP to spawn a pane per agent, drive each one, and aggregate. Activates when the user asks to "run agents in parallel", "compare X agents on Y task", or "fan out to multiple agents".
zhitongblog/unterm · ★ 19 · AI & Automation · score 69
Install: claude install-skill zhitongblog/unterm
# parallel-agents You're the *outer* agent. The user wants you to fan a task out across multiple *inner* AI coding agents (Claude Code, Codex CLI, Gemini CLI, OpenCode, Aider), each running in its own Unterm pane, then bring the results back together. ## Why this works on Unterm - `session.split` carves panes from outside — one pane per inner agent. - `unterm-cli agent launch <id>` (or the MCP equivalent) starts the inner agent in a pane with the right cwd + auto-wired MCP back to Unterm. - `screen.scrollback_text` reads each inner agent's output as a single string — feed it back to yourself to decide what to do next. - Every pane is independent (own PTY, own auth profile, own shell history) so the agents don't trip over each other. ## The pattern ``` ┌────────────────────┐ │ outer (you) │ │ · plans the fan │ │ · reads results │ │ · reconciles │ └────────┬───────────┘ │ Unterm MCP ┌───────────────┼────────────────┐ │ │ │ pane 1 (claude) pane 2 (codex) pane 3 (gemini) gets task A gets task A gets task A runs its loop runs its loop runs its loop │ │ │ └───────────────┼────────────────┘ ▼