Kevin Simback 🍷

Are agents eating up your usage limits? If yes, then you should be incorporating decision models into your routines Let’s reduce your routine’s token consumption by 90%+ and run more frequent routines Allow me to explain One of the most useful things agents do is “wake up” and check on things for you, then take action depending on what it finds This is what makes them proactive and feel like they’re working for you, not just a tool you use We used to call them cron jobs, but someone smart rebranded them to routines - much better The frequency that your agent wakes up and runs a routine is called the “heartbeat” and every time it wakes up, it has to read a bunch of context and possibly call tools - this uses a lot of tokens In all the instances when it wakes up and doesn’t find anything that needs action, it wastes all those tokens Ask your agents to estimate how many tokens are used on each run of your routines - it’s probably 15-40k and if you run a lot of routines frequently it really adds up So when I see people complaining about hitting usage limits, a big reason is likely in the routines and tokens consumed as part of the wake up process I have a ton of routines in my setup and to help prevent hitting limits, I’ve been exploring use of much cheaper decision models like Jev to check status and only wake the expensive LLM agent model if something actually happened The difference is more like 1-5k tokens of a super cheap model vs 15-40k tokens of an expensive model Added up over the course of a month and it can make a big difference Save the valuable usage for when something needs to be acted on, not on the checking to see if action is needed Ok how to do this? I won’t detail it out because it’s easier to just feed this post to your agent and have it work out a plan for your setup In Hermes it’s easy because the cron already has the hook you need, so a job can run a script before the agent, and that script decides whether the expensive agent wakes up In Grok Bot it’s a little more complicated but still doable Grok Bot doesn't give you that pre-run hook, so the “watcher” decision model has to live outside it which can be a small cron script or something like a Cloudflare Worker. It runs the same hash-then-Jev check and only contacts the bot when it's time to act The great thing about this setup is you can run checks much more frequently without worrying much about the cost - run a routine every minute if you want for near real time monitoring Curious to hear if anyone has tried other methods to optimize their setups

I feel sorry for $GLXY They didn't get rated as an AI play on the way up, but they're catching it on the way down

The AI infra trade is getting unwound in full force right now Don't try to be a hero, let the dust settle

If you’re good at AI video generation, I bet you could easily make $1m+ a year It’s one of the best AI skillsets you can learn right now Why? It’s still pretty hard to do quality video I’ve produced some ok videos but nothing good enough, it’s a skill issue for sure But there’s so many new tools now, this being one of the latest Once the tools become cheaper, the skills will diffuse so your window to become an expert at this is maybe 6-12 months

Compute markets (GPU rental prices) seem like an obvious winner here Less new AI datacenter buildout = more competition for existing compute That's likely why we see the hyperscalers up big today whereas the AI infra trade getting hammered

Meant to say GPU prices

Guy who runs his entire company on Grok Bot is now losing it after being “on pace” to spend $4k on inference I have a number of things to say on this: 1. Inference is still highly subsidized, so if you’re maxing out sub plans you’re getting way more $ value relative to cost 2. There are plenty of ways to optimize, he’s taken a number of measures but you can’t expect to run a team of 24 agents and not use up a lot of tokens 3. If you’re getting a lot of value for $4k a month, which is about half the cost of hiring a capable human who can likely do a fraction of that work, then consider it well spent 4. If the unpredictability is the issue, set limits and monitor them daily, simple 5. There are many low cost alternatives, if the $4k to “run your business” is really eating at you then migrate to a local AI solution or a different agent harness I’m ready for whatever hate comes at me in the comments

“Most companies don’t need frontier AI” Here’s the problem with that line of thinking Looking at the company through a vacuum, I agree - you don’t need frontier models to automate most business workflows The cheaper models are good enough, and you can post-train your own for specific workflows But markets don’t operate in a vacuum The threat is that a new competitor will use frontier AI to reinvent your business model and eat your lunch The AI-native models don't need to worry about automating legacy workflows or connecting old inefficient systems together The flaw is an assumption that their business model is safe, and that the goal with AI is to just become more efficient So the need for frontier AI is not to boost productivity, it’s to disrupt your own business model before someone does it for you

This is the most annoying part of X It’s just hyper sensitive to any topic you engage with, even momentarily, and then feels like all my prior FYP content is mostly gone Can we dial this back please?

I can feel this guy’s pain from reading the post At a small company, the CEO needs to own AI strategy, or be smart enough to hire someone who does, but most can’t afford that The worst is a CEO who throws AI slop over the wall and creates chaos from it

What should we use for a company brain? I get that question a lot, especially since joining @codos_ai given that's our core product The answer for most small companies is pretty simple and this article lays it all out If you've built a great tool or have a killer approach, I'd love to hear about it

To do this properly and get the same benefits @Jason mentions, you need: 1. CEO/leader who is fully pilled and wants to change things up, be ahead of the curve 2. An org that is fully bought in, willing to experiment, learn to push boundaries, get uncomfortable 3. Ability to move incredibly fast to test out new tools, methods and workflows, quickly discard what doesn't work and keep pushing on whatever does work 4. Willingness to spend on tokens without needing decimal point precision on exactly where ROI will come from, just a belief that it will come This rules out most organizations, but for those rare exceptions it is game on!

I fully expect we get rugged over time on generous sub plans, but we should keep using them while they last But there's no excuse for getting rugged on your context - this is what you need to own Plan prices will move, but if your context and workflows are portable, switching models or plans is mostly a config change

Waking up to agents that did work for me over the night is still something that amazes me I’m sure at some point we’ll all take this for granted but let’s just reflect for a moment on how incredible this is 🙏🙏🙏

As much as I love Elon and think Grok Bot is a great product, this is a stark reminder why you need to own your intelligence

That was a close one! My shoe size was theoretically accessible to anyone for a full minute But seriously, treat your agents like you would people - delegate responsibly

Wow, can’t wait to see what November brings

100% of companies will need a shared context layer (company brain) 99% of companies don’t know this yet 1% can create edge from this right now

If you want real impact from AI, an incremental approach won't get you there Set a goal you can't hit without it, then go bolder!