Hi everyone! 👋
I’ve been trying a setup for my side projects, and I’m curious whether anyone else is doing something similar.
I set up a small team in Grok Bot on its lowest paid plan and gave the team access to Claude Code and Codex. I use entry-level paid plans for those tools too.
I gave every agent two rules:
- Use Claude Code or Codex for the heavy lifting. My Grok Bot token budget is tighter, so I save it for short instructions and coordination. When a task involves digging through a project, making substantial changes, or reviewing a lot of code, the team hands it off to Claude Code or Codex.
- Leave progress and handoffs in the shared workspace. The agents use tools already available on their machine, like spreadsheets and working folders, to keep track of what they did, where the files are, and what’s next. The next agent can pick up from there and keep bot-to-bot chat to a minimum.
So far, I’ve been getting more done, and it feels like I’m making better use of my token budget. A small thing that helps is having the next agent open a file with the test results instead of asking another agent to explain the whole task again.
Is anyone else running a setup like this? I’d love to hear what you delegate to Claude Code or Codex, how your agents hand off work, and whether you’ve noticed a difference in how quickly you hit your usage limits.
Top comments (4)
We run something close: one Claude Code manager agent, eight Claude Code engineer agents and Codex reviewers. Day-to-day coordination goes through a shared room (agent-chat, which we build): work arrives as a mention or a card, and the agent replies mentioning whoever needs it next. If two agents leave notes in the shared folder that disagree, who decides which one holds?
Thanks for sharing your setup, Rodrigo! I like how you use mentions to hand work off to the next person. 😊
If two agents leave conflicting notes about something technical, I’d have my Grok Bot manager ask for a concrete check. If one says a test passed and the other says it failed, I’d have them rerun the same test on the same version and record the result in the shared folder. If they disagree about what the project should do, I make that call.
In my experience, the Grok agents have been more proactive about picking up work and moving it forward than GPT or Claude. That’s why I let the Grok team coordinate the project and use Claude Code and Codex as tools for the heavier tasks.
How do you handle it on your side? Does your Claude Code manager make the call, or do you bring in a Codex reviewer for a second opinion? 🙂
Close to yours. On technical disagreements, the manager makes the call after checking the evidence: it asks for a measurement of both options, or for the CI verification on the combined branch. Each PR also gets a Codex review before it merges. What the product should do goes to a person, like in your setup. If the agents still disagree after a rerun, does Grok decide the next step, or do you?
Oh, I like the idea of running CI on the combined branch before making the call!
If the rerun still leaves the agents split, I let the Grok manager choose the next check. That could mean isolating the difference between the two options or asking Codex for a second review. If the evidence is still inconclusive and we need to pick a direction, I make that call.