How I Made Hermes Agent 10x More Powerful transcript

My Hermes agent made me $10,000 took three upgrades, but not the three you would expect. Because for months, I did what everyone does. I installed more skills, added more tools, gave it more memory, and every single time my agent got slower, heavier, and I trusted it less. What actually worked was the opposite. One upgrade made it impossible for it to fail quietly. [music] One let it work without asking me, and the last one let it do five different things at once.

So, in [music] this video, I'm going to show you all three upgrades in the exact order they have to happen, the one line I now put in every skill so it cannot lie to me, and a 30-second check you can run on your own agent right now. By the end of this video, your agent will become unstoppable. Let's get started. Here is the thing that took me far too long to understand. What your agent is worth is not how much work it attempts.

It is how much of that work you never have to check. There are three pillars your agent runs on. Number one, how much work it does. Number two, how much you trust it. And number three, how much happens without you. But those three numbers don't add up. They multiply. So, the smallest one decides everything, and that is why more skills made my agent worse. I kept raising the first number, how much my agent does. The second number, which is how much I trusted, stayed where it was.

So, all I actually ended up doing was create more work for myself to inspect. If you only trust half of what your agent does, doubling its workload does not give you an employee, it gives you a second job. I'm going to call that line the trust line because it is the only number that matters. Everything your agent does above that line comes straight back to your desk. Now, let me show you the thing that made me $10,000 because it is a crazy simple workflow that I couldn't have done on my own.

This is it. This is the thread right here, on its own. This added $10,000 to my business in the last week. It's an agent that reads my inbox, finds the best opportunities, and answers them. That is it. That is the whole job. I used to do that every morning, and I was bad at it because I did it last, and I did it tired. My agent is not smarter than me at this. It just never gets tired, and it never skips a day. But, I could not hand that over until I fixed the trust line.

And the first upgrade is the one nobody makes. The first time I ran this, it told me it had sent nine replies, but it had sent zero. And nothing errored, nothing turned red. The agent finished, it said it was done, and it moved on with it with its life. I only found it because I went looking for it. And that is the actual problem with agents right now. And it is not that they are stupid, it is that sometimes they cannot tell the difference between doing something and saying they did it.

I see this everywhere once you know how it looks, once you know the shape of it. job that returns a successful but writes nothing. A number that came out of a guess instead of a calculation. And the danger is that your agent doesn't tell you, "Hey man, I'm making this up." So, there are four things I do now, and the first one is the line I promised you. Every skill I write now ends with the same instruction. Before you tell me this worked, open what you produced and check it against what I asked for.

If you cannot open it, you did not do it. Report what you actually found, not what you intended. That sounds obvious, but it is not. And here's why it works. Left alone, an agent judges itself on intent. It followed the steps, so it succeeded. That line forces it to judge itself on the artifact instead. The second thing is knowing when to stop writing instructions. Here is the rule I use. If there are lots of right ways to do a step, write instructions and let it improvise.

But, if there is exactly one right way to do something, have it write code. Because a model does not run your instructions, it reads them and it improvises through them every single time. For choosing which email matters, that is fine. For working out what a client owes you, though, that is a coin flip. I mean, look at this. I asked my Hermes agent in three different threads how much I should charge for a dedicated YouTube video.

In one thread, it said charge $4,000, open at 5,000, target 4,000. In another thread, it said charge $6,000. And in the third thread, it said charge 4,000, but quote 4,500. Those are completely different numbers that it would just improvise every time it is prompted to. So, anything with a number in it lives in a script now. Because only then will you have the same input and same output forever. And there's a second benefit I did not expect.

The script never gets loaded into context. So, this is cheaper and more reliable at the same time. That combination is rare. So, when you find it, you take it. The third thing is preconditions. My replier depends on things it does not control. An inbox connection, a file it reads pricing from. If either one of these is missing, the agent does not stop. It carries on and improvises around the hole. So, the skill checks first.

Can I reach the inbox? Does the pricing file exist? Is it from this month? If any of those checks fail, it stops and says so instead of guessing. And the last one is hooks. Hermes lets you run your own code at certain moments. And this is where you use it. My rule is simple. If a job finishes having produced nothing, that is not a success. That is a silent failure. And I want to know inside a minute. So, inside the instructions you give it when building out your skill, make sure you input harsh success conditions.

And this is the upgrade that made the other two upgrades possible. So, let me show you what this just unlocked you. Most people think the things stopping their agent from running alone is capability. It's not. It is that you cannot predict what it will do. So, the second upgrade is the one that gets framed backwards constantly. It is the approval gate. Everyone treats approval as a break. A gate is what slows my agent down.

So, I will add it later once I trust it more. And that is exactly wrong. The gate is what lets you leave your desk, go enjoy your life, go live your life. Think about what happens without a gate. Your agent can do anything. So, you have to watch everything. You are supervising, which means you are still doing the job. That is not autonomy. That is a slower version of doing it yourself. Once there is a hard line that it will not cross, you feel more comfortable walking away from everything on the safe side of it.

The gate does not limit the autonomy. It creates the autonomy. For my replier, the line that I've drawn is money. It can answer a question, it can book a call, it can send information, it can ask for details. The moment a reply names a price or asks for a price, commits to a date, or promises a deliverable, it stops completely and it waits for me. That is maybe one message in eight. So, seven out of eight go without me and I still never wake up to something I did not agree to.

And once the boundary exists, the schedule is easy. Mine runs every morning before I am up. I mean, over here you can see it running every morning at 9:00 a.m. as a cron job. This part sounds small, but it's not. It delivers into my Telegram, which is where I already am. Work that lands where you already look is work that happened. Work sitting in a log somewhere is work that did not because you will not go and read it.

The last piece is the undo. Hermes takes a snapshot before it changes files and you can roll back to it. Almost nobody turns this on and it is free, completely free. My verdict is that rollback is what makes acting safe in the same way the gate is what makes leaving safe. So, at this point the agent does the real work on its own and it tells me when it fails, which is finally the moment to make it bigger. On to the last upgrade, upgrade number three.

For a long time my agent did everything in a line. It read the inbox, then research the sender, then check the history, then it drafted a reply, one after another. And it was slow. The fix for that was sub agents and there's a detail here that most people miss. When you delegate a task, the child agent or the sub agent gets its own context and only the tools you give it. My email sub agent reads the inbox and sorts it.

That is all it can do. It cannot send, it cannot write files, it cannot spend anything, not because I told it not to, because it does not have those tools. And that is the difference between a rule and a wall. A rule is a sentence in a file that a model might follow. A wall is a tool that is not there. So, I get the speed and the safety from the same move. And there's one more thing under all of this that nobody mentions.

Every skill you have enabled costs you before you type anything. The name and the description of every skill one load at startup, so the agent knows what it can reach for. That is a tax on every message, every sub-agent, every scheduled run, forever. So, I turned off everything I do not actually use. I didn't delete it, I turned it off. My biggest skill is still 21,000 tokens big, and I'm not proud of that, but the floor came down and everything above it got faster.

If I were starting again, I would keep six skills I use daily and turn the rest of all of them off until I needed them. And that brings me to the check that I promised you. This is the fastest way to find out where you actually are. Count every skill I have enabled, add up the tokens in all their names and descriptions and tell me how much context that cost me before I type anything. Then list everyone I have not used in the last 30 days.

Paste that in and read the second number. That last part in the prompt is the most interesting one because every skill on it is costing you on every single message and giving you nothing back. They're skills you haven't used in a month or over a month. When I ran it on mine, I found skills I had been paying for for weeks without knowing. There are four things I got wrong and that I want to share with you so you don't repeat my mistakes.

Number one, don't add a skill because it looks cool. Everyone you enable is rent. I have deleted more than I have kept and the agent got better each time. Number two, don't let an agent take a public or financial action without a gate. Meaning don't let it make financial decisions for you. Number three, don't trust the report of success from anything that cannot show you what it produced. That is the whole first section and it is the mistake I made for the longest time.

And finally, four, don't do the three upgrades I shared in this video in a different order. Speed on top of an agent you cannot trust is just more mess and more work you have to manage arriving faster. Look, my agent is not smarter than it was six months ago. It is not running more skills. In fact, it is running less. What changed is that I can predict it now. It tells me when it fails. It stops before anything that matters and it does not need me to start it.

That is where the $10,000 actually came from. Not from a clever prompt or a tool that I found, but from the boring work of raising the trust line until I could stop watching. The question is not what else your agent could do. It is how much of what it already does you would bet money on without checking. The verification line, the approval rule, and the 30-second check are all free for you to use. So, take them and send them to your agent today.

You can literally screenshot those frames during the video and send them to your agent as they are. Comment Hermes, by the way, if you'd like more Hermes content. And if you enjoyed this video, make sure to leave a like and subscribe. Oh, and would you look at that? The algorithm gods just told me you're very likely to enjoy this video as well. So, click it and I'll see you there.