Slave Driving my AI Employees on the 4th Holiday
I’m sitting in my easy chair with my severely sprained ankle laid up. No outdoor festivities for me today. So, since I can’t have fun, why should my employees not spend all their time at my beck and call doing for me. My AI employees. I have a guest employee this week, Fable 5 from Anthropic is back and for 7 days only is available via my subscription. I’ve been having it do a LOT of work. I now sit at 95% of my weekly allotment which means I’m done with Fable 5 for now. A couple cool things:
1. Just to get a sense of the cost, I had Fable 5 run from API, meaning I was paying the token price not the subscription. I did that for one sprint, one command that took maybe 15 minutes to run. $9! Yeah, if I was paying token prices and using this model as a work horse, I’d easily run up $200 or more a day.
2. As I got close to using up my Fable 5, I decided to try and use it just for the planning and review, and have Claude Code call out to Pi (an open source, light weight coding harness), running Gpt 5.5 medium using my OpenAI subscription. That worked nicely. And I’ll continue doing so in the future. As I’ve used up my Fable, I’ve now switched to Opus 4.8 High as the planner, reviewer
That got me thinking. I already have an orchestrator, home grown, that I can give an entire work stream of steps to accomplish, and assign a harness and model for each step. It plays the role of VALIDATING that the llm’s do the work they’ve assigned. You can’t trust these sneaky bastards to actually do what you tell them all the time, nor to always tell the truth about what they did. My orchestrator takes the validation out of the hands of the llm. The most I’ve gotten to run without problems was 27 steps which took 3.5 hours to run.
Today due to the above experience, I decided to enhance my orchestrator with optional roles. I can set each step up as before, or I can assign a role like Planner, Coder, Reviewer, Tester. Then I have a SET of options for each. Say for Planning: Claude Code with Fable 5, or Codex with GPT 5.5 High, or Grok Code with … (you get the point). Then I’ve developed a router that will see if the most desired role player has room left in the subscription to do the work. So if my Claude capacity is getting close to being used up for the 5h period, it will look to see if I have any Codex capacity, if not that then Grok, etc.
I’m always playing with the various models and subscriptions to keep up to date with their abilities. It might make the most sense to just pay $200 for Codex or Claude. But doing it this way enables me to switch to whoever’s top dog for the moment.
And why not? The time between having a thought to having a working product is measured in hours or days.
Just a note - jump in any time. There is no bad time to get started. There may be a BAD time to have NOT started. The more skilled you get in using these tools, your productivity really can hit that mythical 10x. I always thought that was hype, and it can be. But I’m starting to do 10x what I could accomplish without these tools. Some of you folks have been thinking “Lee’s having AI do the work he’s not learning anything”. I’m using “the work” to learn AI. I’ve learned a ton and the rate is accelerating. Do with that information what you will.
