I came across this post and it definitely was a scroll stopper.

https://x.com/coreyhainesco/status/2075921300749799759

I was impressed and felt significant fomo. He was squeezing $8k out of his $200 plan and reminded me that this won't last long.

I remember when it used to cost $5 to get anywhere in SF when Uber and Lyft first were around and now it costs 3-6x that depending on traffic.

We will probably see the same thing play out in 12-18 months with frontier models.

So if two of the best capitalized companies in history are competing to sell me a dollar for two cents I best be token maxxing before they go public and need to please the markets.

Anyone know the German word for fomo token maxxing edition :)

https://x.com/msg/status/2076327946412016095

Rather than stressing, I wanted to do something about the feeling that I was just throwing away the opportunity to be building more software for relatively free while I still can.

Measure before you Maxx

Before getting started, I wanted to take stock of what tokens I have access to because, as our elders used to say, you cannot max what you cannot measure.

I installed Peter Steinberger's CodexBar. It's a macOS menu bar app that keeps every provider's usage limits visible and shows when it resets.

I then made sure that I had all the latest CLIs installed on my machine. I already had Claude ($200), Codex ($200), Cursor ($20) and Droid ($20). I was missing Grok Build (via X premium) and Antigravity (~$150).

Now that I know what I'm working with, I need to make sure my agents do too. These tokens aren't going to drain themselves.

Fable The Delegator

Since Fable came out, the meta has been to use it as a delegator and have Codex be the workhorse, but there are tasks that the new Grok 4.5 can do, or that Composer 2.5 could be better for the task. Evals and routing require more research, but in the meanwhile we are just setting up the round robin for the various harnesses.

So Claude and I stood up fable-delegator. It allows me to plan with Claude and then implement via Codex, Grok, Cursor, Antigravity, or Droid CLIs depending on their own token pools. You can see the claude.md file here.

Fable plans and delegates and currently verifies the work (going to probably adjust verification with a separate agent in the future). The implementation itself gets farmed to whichever CLI has quota. For large projects it runs a check via the CodexBar CLI. If one rate limits, a model-overlap map fails over to another CLI carrying the same model family. CodexBar tells me which pool has room.

fable-delegator has instruction skills for how to use any CLI and what responses to expect. You can see the ones I've set up here.

Of note: certain tasks don't need to spin up a separate harness, so be sure to set up CLIProxyAPI. It's a local server that re-exposes the same CLI subscription logins as plain OpenAI/Anthropic-compatible HTTP endpoints, with no harness. For bulk raw generation: draft/rewrite/summarize/classify/extract at volume, LLM-as-judge, fan-out.

Another way to token maxx is to set up an agent multiplexer.

Herdr: one terminal for the whole herd

My friend Thanh Pham told me to look at Herdr to help orchestrate the various agents and subagents.

Herdr is a terminal multiplexer built for coding agents. You have panes and tabs, and sessions persist after you are done. You are also able to view an agent status glance that shows which agent is blocked, working, done, or idle.

The best part of Herdr is you can have your agent orchestrate the various subagent panes through a JSON socket API, so an agent can drive the whole workspace, not just sit in one pane. You can also have several projects in one place. You watch the work being done, or you let an agent watch it for you and check in from your phone.

I watched two Herdr video walkthroughs at 2x to know what to expect: the first and the second. It was easy enough to set up but I still need to put it through the ringer.

One neat bonus thing I discovered using Herdr was that when Codex CLIs are spun up, you can view them in the ChatGPT iPhone app under the remote tab.

Stay Tuned

I will share more of the things I'm spending my tokens on and will report back in 30 days with how much I was able to arb.

So follow @msg

In the meanwhile, may your token maxxing bring joy and prosperity to the world.

Postscript

While I was writing this, I found someone who had the same idea and built it into their Codex workflow.

CJ Zafir's Codex-Orchestration is a Codex plugin: bring any model inside Codex, give it a role like advisor, executor, or reviewer, and run a workflow where the root model plans and verifies while specialized models implement slices in parallel. He reports 40% fewer limit hits and twice the implementation speed.