My AI agents now pick the right model for every job. They use a cheap model for easy work and an expensive model only when the work is hard. Most people [music] run every task on their most expensive model and that is where the money goes. Here’s the thing, most of what your agents do all [music] day is not writing, it’s deciding. Which model should do this task? Which agent should do it? Is this email important or can it wait? For each of those small decisions, you pay a big expensive model to write a full answer. The model that makes these decisions for me is a new AI model called Jeff. It comes from a company called Type Safe and one of its founders helped create ChatGPT. When he announced Jeff, his post got over 40 million views and they say Jeff is up to 200 times faster and 400 times cheaper than the big models you use today. In this video, I’ll show you how Jeff works, then I’ll show you five ways to use it with your own agents and the last one is the one that saves me the most money. And at the end, I’ll show you how to connect it with two prompts to leverage the full potential of Jeff. Now, when Jeff came out, I was so excited. I wanted it to build me a whole machine country dashboard. Then I found out the hard way, Jeff does not write, not a single word. So, what does Jeff do? You give Jeff some text and a question and Jeff gives you back a decision. But you can’t ask Jeff anything like the way you do with ChatGPT or Claude. Every question must be one of three types: choice, score, or null. Let me show you how that works with my own inbox. This is use case number one. Someone sends me this email. Hi Larry, I want to buy your machine country auto installer today. Does it work on a Mac? Now, without Jeff, this what happens. My personal assistant agents running on their big model [music] like GPT or Claude reads the email and works out what it’s about. Then it passes the email to my developer agents because that agent knows the technical answer. So, two agents read the same email and I pay an expensive model twice. Now, with Jeff, it works differently. Jeff sits in [music] the middle. Before an agent reads my email, my dashboard sends it to Jeff and asks Jeff three short questions, one for each type. Now, first, a choice question. Where does this email belong? Sponsorships, customer support, invoices, or newsletters. Jeff picks customer support. Now, in my dashboard, customer support emails go directly to my developer agents. So, the email goes straight there, and my personal assistant agents never read it. Second [music], a score question. How urgent is this email from 1 to 10? Jeff says 10 because this person wants to buy [music] today. So, the email is marked urgent and goes to the top of my list. Ted, a new question. Is this person waiting for an answer? Jeff says yes with 95% certainty. So, the email is marked needs a reply [music], and my developer agents writes a draft right away. I check it, and I send it. So, only one agent reads this email, the one who can actually answer it. And what about the newsletter? Jeff puts it directly into the newsletter category with low agency and nobody [music] waiting. So, no agents reads it at all. Now, that’s how Jeff saves me money, so I can save [music] time. Every email gets sorted out in a split second for a tiny part of a cent. And my big models now only read emails they really [music] need to answer. Now, you might ask, how good is Jeff really at this? I wanted to know that, too. So, I tested Jeff. I gave it 100 emails from my own inbox to sort into four groups: sponsorship, customer support, invoices, and newsletters. Then I went through the results one by one. Jeff put about 90 of 100 emails in the right group, and the best part is it sorted all in a few seconds. The whole test cost me less than 1 cent. So, Jeff works. But you might be thinking, my big AI model can already do all of this, and that’s true [music]. But a big AI model is slow and expensive for small decisions.
It reads every word, it writes long answers, and its answer can change from one run to the next. Jeff gives you one short, clear answer in a split of a second for a tiny amount. You will see the difference in each of the next four use cases. Use number two. This is a new question. No more unnecessary messages from your scheduled jobs. Say your agent checks your email every hour. Normally, your big model writes you an update each time, even when there is nothing new. A message like, “I checked your email, nothing is important.” Now, with Jeff, the job first asks Jeff, “Is there something here that Mona needs to see?” If Jeff says no, the agent stays quiet. That saves you tokens because your big model no longer writes 24 empty updates a day, for instance. And when your phone does buzz, you know it matters. [music] Jeff is as useful when you also have a long list, bringing us to use case number three. This is a score question. Say you have 100 leads from your website. You want to know who to call first. Normally, an expensive model reads and judges every single one. With Jeff, every lead gets a score from 0 [music] to 10 for how ready the person is to buy. Your agent sorts the list by these scores, and the hottest [music] leads end up at the top. That saves you money because the whole list costs one or two cents, and it saves you hours of reading. Now, once you know what matters, which agent should work on it? Bringing us to use number [music] four. This is a choice question, and the options are your agents. Say you tell your main agent, “Find three competitors and compare their prices.” Normally, an expensive model has to think about which of your agents should do it. Now, with Jeff, your main agent just asks Jeff. Jeff just picks your research agent, and the task go there. If Jeff is unsure, your main agent asks you [music] first. That saves you tokens because your expensive model never spends them on picking who does the work, and the task [music] starts very quickly because Jeff understands in under a second. Honestly, when I tested this out myself, I didn’t see any difference with messages going directly to one agent versus messaging going through [music] Jeff to one agent because Jeff responds in a fraction of a second. Each of the four use cases works on one job. The last one works on every task your agents do every day. Use case number [music] five, the one that saves me the most money. This is a choice question with two options, light or strong. You stop paying your strongest model for easy work. In my Accordo dashboard, every agent has two models. A light one which is cheap, and a strong one which cost me more. Normally, every task runs on a strong one. So, you pay the top price even for a quick summary. With Jeff, my dashboard ask before every task, light or strong? A quick search goes to light. Code, a long plan, or anything about money or clients, for instance, goes to a stronger model. If Jeff is unsure, it just goes with a stronger model. So, a hard [music] task never lands on a cheaper model, for instance. Same agents, same memory, only the model changes. Most of my daily work now runs on a cheaper model, and that is where [music] the biggest savings come from. And it cost no time because Jeff chooses in a fraction of a second. Now that you know how Jeff works and five ways to use it, let me show you how you can add Jeff to your own agents. It’s two [music] steps. Get a key and paste the setup prompts. Step one, where to find an API key to add Jeff to your setup. There are two ways, just pick one. The easiest is straight from Type Safe, the company that makes Jeff. It’s free to start. Go to console.typesafe.ai and create an account. Type Safe gives you $5 of free credits. Now with Jeff, that goes a long way because Jeff only charges for the text you send it, and its answers or outputs are free. $5 covers about 120 million tokens [music]. And once you’re on the console, open the API keys, click create a key, give it a name, and copy that key right away because Type Safe shows it once.
Keep it in the notes for a minute. The second way is through Open Router. Pick it if you already use Open Router with your setup because of course, with Open Router, you have one API key that gives you access to a bunch of models. On Open Router, just create an account, open credits, and add, let’s say, $5. Then open API keys, click new key, and give this key a name, and add [music] a small credit limit, so it never spends more than you allow. When you’re done, just copy the key. Now in step two, open the description of this video, copy the prompts one by one, and paste them in your normal [music] chat with your agents. The first prompt connects your key, and one rule here, never paste a key into a chat. Instead, your agent gives you one command. You run it in your terminal and paste the key in there. [music] So, it goes straight into a safe file on your server and never shows up in your chat [music]. Now, your agents can connect and talk to Jeff. The second prompt makes your agents build a skill called Jeff Decides. A skill is a saved set of instructions that your agents can use again and again. [music] Now, with this skill, any agent in your team can ask Jeff a choice, score, or no question. So, that’s it about Jeff. Most AI model writes Jeff Decides and every answer comes from a number that tells [music] you how sure it is. Now, once you start to see your work as choices, scores, and yes or no questions, you’ll find use cases for Jeff everywhere. Tell me in the comment sections which use case you would try first. So, if this video helped you, hey, hit the thumbs up and subscribe so you don’t miss the next one. Thanks for watching and I’ll see you in the next one. Bye.