I burned a billion AI tokens. The fix ran while I slept.

Dave: This week, I burned over a billion, a billion with a B, tokens on ChatGPT through my Hermes Dave: agents doing some software development, and here's what I learned. Dave: Welcome to the AI Operators Field Notes. Dave: This is a new show segment where we're just going to give, like we do on the banter live Dave: show, just our real points of view and opinions on what we're doing with these AI tools and Dave: some of the lessons we're learning. Dave: So, if you are using Hermes and you're keen to use it for software development and you Dave: want to know more about these billion tokens that I burned, and it's interesting here, Dave: right, because these are Luna tokens, primarily Luna that I was using in max mode, because Dave: the Luna pricing is so cheap, I was like, can I get away using Luna max for doing most Dave: of my work? Dave: So, you can see a lot of these were Luna tokens with a little bit of terror because my defaults Dave: just flipped back into terror all the time, and I did use a bit of Sol for Dave: oversight just to try and make sure the Luna were running on course and needed some Dave: course correction. Dave: This doesn't count all the Grok tokens that I've spent this week as well. Dave: So, why am I spending Grok tokens as well? Dave: Well, that's because I've hit my usage limit. Dave: I have 0 % of my GPT. Dave: In fact, that billion tokens is across two ChatGPT business accounts that I have access Dave: to that I'm using all of the Codex quota from. Dave: This is one of the things with Hermes. Dave: You can actually authorize multiple different accounts, even with the same provider, and Dave: then set up a whole bunch of rotating policies. Dave: Do you want to let one fill up and then use the next one? Dave: Do you want to take turns using which plan, which plan, which plan, swapping your subs Dave: together, as well as just supporting all kinds of models? Dave: It's one of the things that really keeps me with Hermes is the fact that I'm not tied Dave: to one model, and when I run out of usage on one model, I can jump to another model. Dave: I can have different plans. Dave: Unless, of course, it's Anthropic, because they say, no external harnesses. Dave: So I've got no time for Claude anymore. Dave: This week, there's been a lot of talk about bots and Grok bots and so on. Dave: Hermes have come out with their own bot thing, but I'd rather just go to the agent profile Dave: that I know I want to talk to that I've already configured and use session -based chats. Dave: I think the bot thing is just a consumerization thing, trying to help people get their head Dave: around it. Dave: But Hermes agent. Dave: Super duper flexible. Dave: In fact, they're actually pushing this. Dave: This is one of their latest tweets. Dave: Use Hermes agent however you want. Dave: Modify it, extend it, share it, fork it, use it locally in the cloud, use desktop app, Dave: Chewy, talk, text, run a Fortune 500 on it, use one model or 15 or blend them with Dave: a mixture of agents, use it to plan an elopement, debug an event, summarize a lease, blah, Dave: blah, blah. Dave: Have it scan, use weekly to settle that stupid bet you made with your brother that one time. Dave: It is an all -in -one, very flexible, very capable. Dave: Harness, but with so much power, there is probably some responsibility, but there's Dave: also a lot of different ways to get to the goal. Dave: Um, and so I've been trying to, I've got a few different SAS ideas that I'm trying Dave: to build at the moment and I'm doing some fairly large software builds. Dave: And that's why I was like, maybe if I can just thrash Luna max, I can get more done, Dave: uh, within my plans without having to buy extra plans and extra tokens and live within Dave: my means. Dave: And so one of the things that, uh, Hermes has had in its skill library for some time Dave: is sub -agent driven development. Dave: So I was like, this is great. Dave: I've, I spent some time doing up all my, my spec of what I wanted to build. Dave: And then I said to Hermes, sub -agent driven development, give it to Luna max. Dave: And it was doing it and it was doing it. Dave: And what was happening apparently. Dave: Um, but then was, I watched the, the, the, the scroll. Dave: I started seeing agent timed out. Dave: Didn't finish its job. Dave: Didn't change it. Dave: Didn't change any files. Dave: Um, I've just done it myself. Dave: And I was like, that's, that's not what I want. Dave: Now, one of the great things about all these agents is you can get them to debug themselves. Dave: So I've actually got a master Hermes agent. Dave: So the way I've got my Hermes fleet set up is I have sort of the, my main Hermes profile Dave: that helps me administer Hermes. Dave: And then I have different agents set up for different projects and tasks. Dave: One for this podcast, one for my fractional consulting, one for each of the different Dave: SaaS applications that I'm building, et cetera, et cetera. Dave: So I said to my master agent, I was like, Hey, can you go look at the logs and work out why these agents are timing out? Dave: And what's going on? Dave: I went, oh yeah, there's some old limits around how many iterations it can do. Dave: We'll fix those up. Dave: But then I was like, okay, we'll try it again. Dave: And then it came back and said more timeouts, more timeouts. Dave: Um, so I got it to dive in deeper. Dave: And this is when I realized that because it was in these early stage of software development, too many of the tasks was sequential. Dave: I had to build this thing before I could build this thing. Dave: And I had to get that in place before we could do it. Dave: We could build the next thing. Dave: These things all built on top of one another and sub -agent development just wasn't a fit. Dave: You couldn't run multiple parallel streams. Dave: And so Lunamax not being super smart because it's Luna was just like farming out work to these agents that couldn't finish it. Dave: Now running in circles and it was a really inefficient use of my tokens. Dave: There's your field time. Dave: If you learn nothing else, take that away. Dave: Uh, so I, that's when I brought Sol in and said to Sol. Dave: Hey, I want to course correct, make sure we're running on the right path, but also like, why isn't this working? Dave: And it sort of helped me dive into this sort of story. Dave: And I said, look, the reason I use sub -agent development is because I wanted to, I've done the work on the spec. Dave: This is what I want to build. Dave: And I just wanted you to build it and just run for a long time and build it. Dave: And I thought you using the sub -agent development skill was a good way of doing that. Dave: Sol reviewed all the Hermes documentation has access to within Hermes. Dave: And said, you know what, you should be using the Kanban board instead. Dave: So here we've got the Kanban board. Dave: This is a screenshot from, uh, Noosa's documentation, but the Kanban board, if you're not familiar with Kanbans is just like a Trello board or something like that. Dave: It's a task board where you can assign tasks to agents. Dave: You can have different agents set up. Dave: You can assign it to different models. Dave: And so what I got Sol to do was break down the project into task cards that were assigned to Luna. Dave: Max. Dave: And so now Sol was really overseeing and it was managing it through rather than this, this random sub -agent development skill was managing it through a more structured form using the Kanban board. Dave: Uh, and that seemed to do a lot better. Dave: Uh, and so I'm a big fan of using the sub -agent driven development when it makes sense, when you have multiple parallel tasks, but if you're kicking something off or you want this to run semi autonomously and have automated checks and reviews, talk to your Hermes agent. Dave: Around. Dave: Setting it up in a Kanban board. Dave: I did use Kanban when I first went to Hermes and I kind of went too far and too automated. Dave: And if you've seen my IRL talk that I did, I just got out of control and there was just too many things getting blocked and failing, and I went too hard and didn't manage it. Dave: But on a specific project with a specific spec and getting an intelligent agent, like intelligent agent, like Sol to do some of the planning and theory and set up and then put it all in the Kanban board that enabled me. Dave: To actually get things done or walk away for a long period of time, go to bed, wake up and actually find the project had been moved forward. Dave: And now Hermes, that multi -model agent thing that you can do, I ran out of all my GPT and I could say, okay, I've got a bit of grok left. Dave: I'll use 4 .5. Dave: That's smart enough. Dave: What I need rock 4 .5. Dave: Look at the documentation. Dave: The Sol wrote, look, we're up to on the Kanban board. Dave: Let's farm out the next set of tasks to composer .com. Dave: 2 .5, which is another cheaper model on the grok plan. Dave: So I don't need to burn grok 4 .5. Dave: I can use the really capable cursor composer, 2 .5 for executing the tasks. Dave: The grok is orchestrating. Dave: And so it's gone through and run a whole bunch of tasks through there. Dave: So there's different ways to skin the cat. Dave: And I'm not saying that this is necessarily the best way, and I'm not going to be hybrid. Dave: You must do it this way and you'll learn 10 X. Dave: And that's not how we roll here on the AI operators. Dave: This is field notes. Dave: This is my personal learnings that I think. Dave: I've learned from, and I think are useful to share and let other people know about. Dave: Now I need a little bumper for this one, but a robot, we've got a hot tip for the people. Dave: If you are using something like codex, specifically codex, and you run out of your usage, Dave: or you're finding you're using all your codex usage really quickly through whether it's the codex harness Dave: or whether it's through something like Hermes. Dave: Did you know, if you just go to chat GPT .com, that doesn't count against your quota. Dave: So if I come back to my little usage screenshot here, usage is shared across codex, work workspace Dave: agents and chat GPT for Excel. Dave: It doesn't include chat conversations. Dave: So one of the things I've started doing, like as much as I'm deep and all in on Hermes Dave: and I love that it's in telegram and I love that it's on my desktop and everywhere. Dave: It's great, but it also uses my tokens. Dave: Whereas if I want to start doing some high level research planning, thinking about what Dave: I want to do. Dave: Framing out the spec for my next build. Dave: I'm now moving back into chat GPT .com when I'm, when I moved in, if you check my substack Dave: as I moved into this sort of more sovereign environment of running models on my desktop Dave: and having all my markdown files and having all my context here and not captive up in Dave: a web platform, I haven't been using chat GPT .com as much, but the benefit of being Dave: able to use the intelligence and format and research and find things out and frame up Dave: specifications. Dave: Basically for free, like, yes, I'm paying the subscription, but it's not coming out Dave: of all my coding harness codex. Dave: That's huge. Dave: So that's my extra little hot tip. Dave: If you are finding you're hitting your codex usage limits really quickly through whichever Dave: harness you're using. Dave: Don't forget chat GPT .com is really good for some of that preliminary research specification Dave: framing up kind of work. Dave: And even if it's like little one page widgets and things use the artifacts in that. Dave: I want to do that. Dave: I wanted to build a little thing for scanning books on a library shelf and rather than build Dave: a whole app. Dave: I just bought a custom GPT. Dave: So sometimes the old ways are still the good ways. Dave: That is ARL Produce Field Notes. Dave: Let me know if you learn something. Dave: Let me know your experience. Dave: Are you just using Hermes as a personal agent getting to help write emails and do some productivity Dave: stuff? Dave: Are you doing serious software development in it? Dave: I definitely am. Dave: I'm not only am I building all of my websites for all my different things, purely in the Dave: Hermes harness now, but also trying to build some larger SaaS applications, which I can't. Dave: We'll talk about in due course. Dave: So please make sure you like and subscribe to the channel. Dave: We'll be back with more content as we always are. Dave: Definitely the live show on Wednesdays, but drop a comment and let me know what your experience Dave: with Hermes has been and we'll see you next time.