Big Idea

The argument in one line.

Hermes Agent's new update turns three separate ideas — multi-model consensus answers, one-command skill creation from any URL, and a git-native coding surface — into a single assistant that gets both cheaper and more capable at the same time.

Who This Is For

Read if. Skip if.

READ IF YOU ARE…

You already run an AI agent tool day-to-day and want to know which new features are actually worth turning on.
You're deciding whether to keep a separate coding tool (like Claude Code) alongside your agent, or consolidate into one.
You want a repeatable way to turn things you read online (tweets, articles) into reusable agent skills without manual prompt engineering.
You're cost-sensitive about running an agent on top of a paid model and want to know what actually drives the bill down.

SKIP IF…

You don't use an AI agent/assistant tool with a skills or memory system — none of this applies without that base product.
You're looking for a neutral comparison across agent tools rather than one creator's walkthrough of a single product's release notes.

TL;DR

The full version, fast.

This update to Hermes Agent adds seven features built around three ideas. Mixture of agents (/moa) sends a prompt to reference models like ChatGPT and DeepSeek, then has the orchestrator model synthesize their answers into one response — useful for complex, open-ended questions. /learn turns any URL, prompt, or completed task into a saved skill or memory automatically, and /journey visualizes everything the agent has learned as a connected graph. Background learning tasks now route to cheaper models instead of the primary (often expensive) model, cutting cost without cutting capability. The desktop app also gained real git controls — diffs, commits, and pull requests — making it a viable full vibe-coding tool, not just a chat interface. The closing recommendation: don't use the new Fable 5 model as your daily driver (too expensive), but spin up a dedicated profile for it and reserve it for genuinely hard build tasks, while using mixture of agents for hard questions.

Free for members

Chat with this breakdown — free.

Sign in and you get 23 free chat messages on us — ask for the hook, quote a framework, find the exact transcript moment, generate a markdown action plan. Bring your own key when you want unlimited.

Create a free account →

Chapters

Where the time goes.

00:00 – 00:23

01 · Cold open

Claims Hermes Agent is the best AI agent made and previews seven new features plus a Fable 5 recommendation.

00:23 – 03:15

02 · Mixture of agents

Explains and demos /moa: prompt goes to reference models (ChatGPT, DeepSeek by default via OpenRouter), aggregator model synthesizes one answer. Live demo generating SaaS ideas.

03:15 – 05:30

03 · /learn — automatic skill creation

Any URL, prompt, or completed task can be turned into a saved skill with /learn. Demos turning a tweet into a skill.

05:30 – 06:26

04 · /journey — visualizing learned skills/memories

A graph view showing every skill and memory the agent has accumulated and how they connect.

06:26 – 07:19

05 · Self-improvement got cheaper

Background learning tasks now delegate to cheaper models instead of the expensive primary model, cutting cost.

07:19 – 09:10

06 · Vibe coding with real git controls

Desktop app now shows diffs, allows commits and pull requests. Demos a vibe-coded Three.js game and a Diablo-style action RPG.

09:10 – 11:08

07 · Recommendation: Fable 5 via a dedicated profile

Advises setting up a separate agent profile just for the Fable 5 model rather than using it as a daily driver; positions it for complex build tasks vs. mixture of agents for complex questions.

11:08 – 11:57

08 · Sign-off / community plug

Plugs a paid community/bootcamp and asks for engagement.

Atomic Insights

Lines worth screenshotting.

Mixture of agents sends one prompt to multiple reference models (default: ChatGPT and DeepSeek) and has an orchestrator model synthesize their answers into a single response.
The default mixture-of-agents combination is ChatGPT plus DeepSeek run through OpenRouter — if it doesn't work, the missing piece is usually an OpenRouter API key.
A one-command /learn on any URL, prompt, or finished task automatically converts it into a reusable skill or memory, no manual prompt-engineering required.
Running /learn on a tweet with a useful tip is enough to generate a new saved skill from it in seconds.
/journey renders every skill and memory the agent has accumulated as a connected graph, so you can see how conversations, skills, and memories link together over time.
Background self-improvement tasks (creating memories, creating skills) now automatically route to cheaper models instead of the expensive primary model, which was previously a major cost driver.
The desktop app now has full git controls inside the chat interface — viewing diffs, making commits, and opening pull requests — making it usable as a standalone coding tool.
The stated reason Hermes is described as ahead of a rival tool (OpenClaw) is specifically this vibe-coding upgrade, not the chat features.
The recommended way to use a new, expensive model (Fable 5) is a dedicated profile reserved for it, not switching your main daily-driver profile.
The suggested division of labor: use mixture of agents for complex questions that need multiple perspectives; use the Fable 5 profile for complex, multi-step build tasks.
Profiles in this tool function as separate agent instances, each with its own model and its own connection (e.g., a Telegram bot) — not just a settings toggle.

Takeaway

Match the AI tool to the job: consensus for questions, one strong model for builds.

WHAT TO LEARN

The most transferable idea here isn't any single feature — it's the decision rule for when to ask multiple models versus commit to one expensive model.

For open-ended, high-stakes questions, running the same prompt through multiple models and synthesizing the answers reduces the risk of any single model's blind spot.
A one-command habit of converting useful things you read (tweets, articles) into saved, reusable instructions compounds faster than re-explaining the same context every time.
Background or maintenance-type AI tasks (summarizing, organizing, remembering) don't need your most expensive model — routing them to a cheaper model can cut cost with no visible quality loss.
When a new, more expensive model becomes available, isolating it to a dedicated use case (rather than replacing your daily driver) avoids paying premium rates for routine work.
Native version control (diffs, commits, pull requests) inside an AI coding surface is what separates a toy chat demo from a tool you can trust with a real codebase.

Glossary

Terms worth knowing.

Mixture of agents: A technique where one prompt is sent to several 'reference' AI models in parallel, and a separate 'aggregator' model combines their answers into a single synthesized response.
Orchestrator / aggregator model: The model responsible for collecting responses from multiple reference models and merging them into one coherent answer.
Reference model: One of the secondary AI models (e.g., ChatGPT, DeepSeek) consulted alongside the primary model when using mixture of agents.
Skill (in an agent tool): A saved, reusable capability or instruction set the agent can invoke later, created either manually or automatically via a learning command.
Agent profile: A separate configured instance of the assistant — its own model choice, memories, and connected channels (e.g., a Telegram bot) — distinct from the user's main/default agent.
Vibe coding: Building software by describing what you want in natural language and letting an AI agent write, and increasingly manage (via git), the code.

Resources

Things they pointed at.

01:40toolOpenRouter

01:40toolChatGPT

01:40toolDeepSeek

11:08productVibe Coding Academy

Quotables

Lines you could clip.

00:00

“Hermes Agent is officially the best AI agent ever made.”

cold-open hyperbole hook→ TikTok hook↗ Tweet quote

09:10

“I would go in, start a new session, and ask your agent to create a new Hermes profile specifically for Fable five.”

concrete actionable tip→ IG reel cold open↗ Tweet quote

10:36

“Fable is good when you want, like, a perfectly built out front end, a perfectly built whatever.”

clear positioning statement→ newsletter pull-quote↗ Tweet quote

The Script

Word for word.

Read-along

Don't just watch it. Burn it in.

See every word as it's spoken — crank it to 2× and still catch all of it. The same dual-channel trick behind Amazon's Kindle + Audible.

00:00Hermes Agent is officially the best AI agent ever made. They just released an update that has several new features that completely change how I use Hermes Agent. In this video, I'll cover the seven new features you need to implement into your workflow today, plus towards the end, show you how you should be using Fable five with your Hermes agent.

00:21This is a must watch. This is a banger. Now let's lock in and get into it.

00:25Alright. I'm starting off with one of the best features ever implemented into Hermes agent, and that is mixture of agents. What is mixture of agents?

00:34Well, let me show you. When you type slash MOA and then you put in a prompt after, it is going to initiate the mixture of agents.

00:43Now mixture of agents? Well, what's gonna happen is my aggregator model, which is my orchestrator, is gonna take that prompt.

00:52So that's whatever the model you're currently using for me, that's Opus four eight. It's going to take that prompt, and it's going to send it to a bunch of reference models. Right?

01:02So we're taking this prompt. It's sending it to a bunch of reference models. These reference models by default are chat GPT as well as deep seek v four through open router.

01:12That's the kind of default way they do it. It takes the prompt, gives it to all of these reference models. The reference models answer the question and then give the response back to the aggregator model, which, again, Opus four eight.

01:26The aggregator model takes all their responses and then synthesizes it into one response, one synthesized answer. According to new research's benchmarks, this results in a much better, higher quality response.

01:42They tested all the models. They tested a bunch of different reference models. Their best combination of models was, again, ChadGBT and DeepSeek, and that is kind of the default in here.

01:52Alright. So all you need to do to make the mixture of agents works is click on the drop down, choose MOA default. That's going to send it both to ChatGPT as well as DeepSeek four through Open Router.

02:04I'm gonna give it a prompt based on what you know about me and my interests, passions, goals, and ambitions. What are five SaaS ideas I can build out today to make money online? That's a prompt you can steal right now.

02:13Feel free to steal that. Use that with your mixture of agents. Hit enter.

02:17Now all the agents are gonna get together, convene, chat with each other, figure out what are the best five SaaS ideas based on all the conversations, all the memories it has about me, and give me a really nice synthesized list of ideas I can start building out immediately. If I wanna turn around after this, go to Hermes and say, hey.

02:35Build that idea out. I really like that a lot. It can do that.

02:38So Chad GBT is talking. Deepsea, Deepsea is talking Chad GBT. They're coming up with different ideas, and here's a nice synthesized answer right here.

02:46And boom. Here's an actually a bunch of awesome really, really good ideas for SaaS. I can take one, start building it out now, and boom.

02:53I have a weekend project for myself. Do the same thing. Set up mixture of agents.

02:57If it doesn't work for you off the rip, it's probably because you do not have open router set up. Open router is default. The second model uses the default first model is ChadGPT.

03:07If it's not working, make sure you have those API keys in. Ask Hermes. It'll set up for you.

03:11Really, really easy to do. Mixture of agents, awesome new feature to get even better answers out of your Hermes agent.

03:18The way I'd use this is just whenever you have kinda more complex questions, it can go many different ways. I use the mixture of agents for it. Makes it really nice to use.

03:27The next major feature in this release is slash learn. This is an amazing way to teach your agent new skills. Basically, the way it works is you can hand your agent any URL, any prompt, or run slash learn after any task, and it will automatically create new skills and memories around what you gave it.

03:48Let me give you a good example here. So if you're anything like me, you're on x all day, you see tweets with cool tips on how to use your age and how to use Claude code or whatever. You're like, oh, I'm gonna go recreate that in my Hermes.

03:58Now what you can do is take those tweets and just do slash learn and have your Hermes make a skill based on every tweet you see that's helpful. I got this tweet here by me. Great tweet.

04:10On coding loops, I'm just gonna take the URL for this tweet. I'm gonna go to my Hermes agent.

04:16I'm gonna do new session. I'm gonna do slash learn, paste in the URL, and I'm gonna hit enter.

04:22And what's going to happen is the agent's gonna go read the tweet and then turn it into a skill. Now to have it explain every single thing you see online, trying to make skills out of it manually for your agent, you just give it a URL, do slash learn, and boom.

04:36You have a brand new skill automatically made. Slash learn is the quickest, easiest way to turn anything into new skills automatically.

04:48Anytime you think of a new process, anytime you see a helpful tweet or article online, you just do slash learn and then type it in after. You can even do it where if you have your agent do a whole long process, you're doing something, you haven't built something, you're like, that worked out great. I want that to be a repeatable skill.

05:05You just do slash learn and hit enter, and it'll take everything from before that and turn it into a skill. It is a really good way to quickly self improve your age. And as you can see here, boom.

05:16New skill just made based on that one tweet. Every time I see an interesting tweet, I do slash learn, put the tweet in after, and it's good to go. Slash learn.

05:24Use the hell out of it. This next one is an awesome one that builds off a slash learn, and it is slash journey.

05:31This is a great way to see everything your agent has learned and all the skills and memories it has. Look at this awesome chart. Now you can see every memory, every skill made from all of your conversations you've ever had with your Hermes agent, and you can see how they're all connected.

05:48You can see how all the skills connect to all the different memories. I started building a game. It showed me all of the memories attached to building games.

05:57Here's my work around building local models and putting them on my computers. You can see all the memories around local models. Really sick way to see what your agent has learned about you as you've worked with it.

06:06Slash journey, you can see this amazing chart, and just a good way to make sure your agent is learning along the way and all the self improvement it's had. Which brings us to the fourth improvement of this new update, which is self improvement got cheaper. Now every time your agent learns something, every time you do slash learn, every time it self improves automatically, it actually automatically switches to cheaper models to implement those learnings.

06:32A big part of the cost with Hermes agents, there's a lot of complaints about cost, especially if you're using anthropic models, is it's really expensive. Well, the issue is is a lot of the things that happen in the background, creating memories, creating skills, learning new things, those are very expensive tasks, and it uses your Opus model if that's what you're using, your ChadGPT model to do those things.

06:54Now it's actually delegating those tasks to cheaper models in order to save a lot of money. So if you're using Hermes agent regularly, you should start seeing much better cost savings as you go along. Major update.

07:05Really awesome to see. Next up is massive improvements to vibe coding. Now when you vibe code inside Hermes desktop, you have full git controls.

07:16You can see code. You can see diffs. You can make commits.

07:19You can open up pull requests. It's all built into Hermes. Now this is an actual viable vibe coding tool.

07:26If you look here, I asked to build a three GS game, which for the record, I'm gonna show you the game in a second. I actually built is incredible. Probably the best three GS game I've ever seen built out of this.

07:35But what you can see here is it creates all the files. You can see all the diffs. You can see all the code changes.

07:41And if we go over here to the right, you can actually make commits and actually open up pull requests straight from Hermes. So this is a viable vibe coding tool.

07:53If you want, you don't need to use Claude code anymore. You can use this and just build out your code and your apps straight from Hermes, and Hermes is a very good vibe coding tool. Check this out.

08:04I just asked to build an action RPG like Diablo, and this is what it came up with. This is actually really, really nice. You can see the graphics are good.

08:12You can see the hit counters. You can see the spells it gave me. It has full levels.

08:16It has leveling up. This is really amazing. This is a really good game that Hermes vibe coded all by itself.

08:24This is probably the most impressive vibe coded three GS game I've ever seen. This is a Opus four eight.

08:30I'm getting really into path of exile two, so this is certainly holding me over.

08:37If I am you, I try vibe coding now inside Hermes agent, especially if you have, like, your ChadGBT subscription plugged in. So you're so you get all those free credits with your subscription. I think you get get away with using this as your vibe coding tool now because you can see the diffs.

08:52You can create pull requests, all of that. Really, really nicely done. This desktop app is incredible.

08:57I mean, this is probably the sole reason Hermes is ahead of OpenClaw right now. So take advantage of it. Test it out.

09:03Do some vibe coding in Hermes. I'm really, really impressed. And the last bonus recommendation I wanna give you here is around fable five.

09:11Fable five is now available in Hermes agent. As you can see here, they have fable five. Here is my recommendation on how to handle this.

09:20I would go in, start a new session, and ask your agent to create a new Hermes profile specifically for fable five. So here's the prompt I would use.

09:30Please make a new Hermes profile called Oracle that uses fable five model from Claude. Let me know when that's set up so I can use it in Telegram. For those who don't know, profiles in Hermes are basically new Hermes agents.

09:42So every profile you set up is a new Hermes agent you can use. I highly recommend setting up a new Hermes profile built specifically for the Fable five model.

09:53And I'm not using this as the daily driver. Not at all. You do not wanna use this daily driver.

09:57It is way too expensive for that. But what I'd use it for is super, super, super complex problems. Right?

10:04You have your mixture of agents option, which is one way to handle complex problems. That I think is more for, like, complex questions and getting really well thought out answers.

10:15What I like to use Fable five for is when I'm building things with Hermes agent. Hey. Build this out for me.

10:20Build that for me. I find coding with Fable is significantly better even when you use mixture of agents. Mixture of agents is good when you need opinions from multiple models to get a nice synthesized answer.

10:31Fable is good when you want, like, a perfectly built out front end, a perfectly built whatever. Fable five will do that the best. So I highly recommend or if you're doing really complex actions across your computers, like if you wanna set up a local model on your DGX Spark that connects to your Mac Studio that does this on another computer, really complex tasks that are multistep across different devices, I would use Fable agent for that as well.

10:58So your line of thinking is complex questions, use mixture of agents. Complex tasks, use your Fable agent. You set that up by setting up a new profile inside Hermes, and it will go and build that out for you.

11:11Then you can set it up in Telegram, and you're good to go. You can go on your Telegram and talk to that agent at any time. Just as a side note, I'm gonna do a full boot camp on Hermes agents.

11:20We have multiple full boot camps on Hermes agent inside the Vibe Coding Academy. Link for that is down below. It's the number one community on AI on the entire Internet.

11:30I promise you will not regret joining. Build alongside a thousand people doing cool stuff with AI. You'll learn a ton of stuff.

11:37Promise it'll be worth it. Link down below for that. You learn anything at all, leave a like down below.

11:41Let me know what you want in the next video. More Hermes agents tips, more Fable five tips, more Claude code tips. Let me know down below.

11:48All my videos are based on your feedback. I hope this was helpful. Make sure to turn on notifications so you do not miss the next video.

11:54I will see you in the next one.

The Hook

The bait, then the rug-pull.

A software update video with a simple premise: seven new features dropped in one release, and the host walks through each one live on screen — ending with a pointed recommendation on how (and how not) to use the newest, most expensive model available inside the tool.

Frameworks