How to make OpenClaw delegate complex tasks to Claude Code & Codex (and aggregate the results)

How to Make OpenClaw Delegate Complex Tasks to Claude Code and Codex as Sub-Agents I’ve been trying to get OpenClaw(/Hermes ai) to handle some really heavy coding tasks lately, but sometimes it hits a roadblock when the task is just too complex for it to do alone. Instead of forcing OpenClaw to do everything, I found that offloading the heavy lifting to specialized CLIs (like Claude Code or Codex) works way better. I set up a workflow using a local broker script that sits between OpenClaw and the CLIs. It runs the CLIs in isolated temporary workspaces so they don't mess up my main project files, and then feeds the clean JSON outputs back to OpenClaw. Essentially, it turns Claude, Codex, and AGY into asynchronous sub-agents for OpenClaw—they just do the work and return a structured JSON report. The best part is that you can just add a simple SKILL.md to OpenClaw, and then you can prompt it like this: │ "Ask Codex and Claude who the next US president will be, and aggregate their reasoning in under 200 words each." OpenClaw will then spin up asynchronous background jobs for both CLIs at the same time, track them, and compare their answers for you. It's a great way to handle tasks that OpenClaw struggles with natively. I packaged the broker script and the SKILL.md into an open-source repo called Relay-agent. If anyone is struggling with complex task delegation or wants to try this multi-worker workflow, you can check it out here: https://github.com/miter37/Relay-agent