The tool that sees where the agent fails
How am I solving my everyday problems with artificial intelligence? I built a tool that helps me manage automations. When I have a communication problem with an agent, the tool figures out where the agent is going wrong, where it cannot solve the problem, where the snag is. That is how I built a system that improves itself. It is called Noesis: it took a lot of time, and today it saves me a lot of it.
Flat rates: lots of tokens, lots of models
Today we have Gemini, ChatGPT, Claude, Grok and every other model. There are so many that they are everywhere. The good news is that we live in the world of flat-rate plans, great for burning through lots of tokens. But to use them well you need many models: if you only use Gemini and ignore the others, you are literally wasting your potential. So I built a system that lets me use all the agents at the same time.
We are the bottleneck
The bottleneck is not AI. It is you, it is me, because we cannot communicate well, not only with a single AI but with all the ones available. We will have more and more compute and more and more computers: the way to use them better is cybernetics. If we connect all these AIs we can build a really powerful workflow. I have wired Grok to ChatGPT, Grokbot, Cursor, the external plugins and Muse. With ChatGPT's voice assistant I can have it make videos and Shorts, check my Airbnb bookings, send messages. An endless range of scenarios has opened up.
The bill: 50-70 euros a month
A small piece of communication is enough to put them together and use the compute these giants are practically giving away. Gemini's basic plan with the terabyte, ChatGPT at 10-20 a month, Muse at 5, Grok at around 25: for 50-70 euros a month you get plenty of models to create content and sort material. But only if you can talk to AIs and use them well.
Linear + ChatGPT: the roadmap the night before
I keep creating projects in Linear, connected to my ChatGPT through its plugin. Linear is a sort of Slack that brings the coding sessions together. So I no longer go by blind attempts like "what do I do tomorrow?". In the evening I set the roadmap in Linear, and the next day I tell ChatGPT: we have the roadmap, today we only execute. The roadmap holds all the connections and plugins, like ChatGPT web's computer use, which works right inside your computer. And when that is not enough, you tell it to use Codex and your CLIs.
Metricool, YouTube and the system that learns
I connected Metricool to my YouTube and to ChatGPT, and asked how my Shorts are doing. The answer: you are doing well, but you need to be more direct with your audience, around 50 years old, and less abstract. Then it gave me the prompts for the next Shorts with that fix already inside. The next Shorts will be better twice over: I will communicate better, and the video generation will be tuned to that audience too. That is the self-learning system. Every day, as I add plugins and connections and fix bugs, the system gets more autonomous, and an autonomous system needs less attention because it starts improving on its own.
Video (in Italian): Il collo di bottiglia non è l'IA, sei tu: collego tutti i modelli e il sistema impara da solo · Italiano: versione italiana · Yesterday: Today's AI agents are the HTC Magic · Community: Body Mind Systems on Skool.