Hermes Agent: Why Self Improving Personal AI Agents Are the Next Workflow
Prefer to watch? Watch on YouTube · Subscribe for new builds · Read the transcript
Hermes Agent from Nous Research matters because it points at the next real shift in AI tools. Not another chat tab. Not another pretty wrapper. A personal AI agent that remembers your work, builds skills from experience, talks through your channels, runs scheduled automations, and gets better the longer it works with you.
That is the part most people miss. The future is not one magic prompt. The future is a system around the model.
When a repo like Hermes shows up with memory, skills, cron, messaging, subagents, and multi model support, it is not just a feature list. It is a signal. Personal AI agents are turning into operating layers for builders.
What is Hermes Agent from Nous Research?
Hermes Agent is an open source personal AI agent project from Nous Research that focuses on persistent memory, skill creation, scheduled automations, messaging channels, subagents, and model choice. The simple version: it is trying to become an agent that grows with you instead of a chatbot that forgets everything after the session ends.
That phrase, grows with you, is the whole story. Most AI agents still act like smart interns with amnesia. You give them a task. They do it. Maybe they do it well. Maybe they hit a wall. Then the session ends and the next time you ask for something similar, you repeat half the context again.
Hermes is aiming at the opposite pattern. It wants the agent to learn from the work, keep useful knowledge, improve skills during use, and search past conversations when the current task needs history.
That is the same direction the entire serious agent market is moving. The model is not enough anymore. The workflow around the model decides whether the agent feels useful every day or random every week.
Why do self improving AI agents matter?
Self improving AI agents matter because the value is not just in one task. The value is in operational memory. If an agent learns how you research, how you write, how you approve changes, and how you ship work, every future task gets cheaper and faster.
Think about it like hiring. A bad hire asks you the same question every week. A good hire learns the process. The first time, you explain the task. The second time, they ask fewer questions. The third time, they just do it.
That is what agent skills are trying to become. They turn repeated work into reusable operating knowledge.
For a builder, that is huge. Your agent should know how to check YouTube performance. How to save drafts to the database. How to verify a source before writing a script. How to avoid mistakes you already paid for. None of that lives inside the raw model weights in a clean, business specific way. It lives in the system around the model.
This is why memory and skill creation are not cute features. They are the compounding layer.
What can a personal AI agent actually do?
A real personal AI agent can research, draft, code, monitor, summarize, schedule, remember, route work across tools, and ask for approval when an action matters. The difference between a chatbot and an agent is that the agent has tools, memory, permissions, workflows, and some kind of long running presence.
A normal chatbot gives answers. A personal agent starts acting like a worker.
- Memory: It remembers preferences, past decisions, files, projects, and recurring workflows.
- Skills: It turns repeated tasks into reusable instructions and scripts.
- Channels: It can work through Telegram, Slack, Discord, email, or other places you already operate.
- Cron and automations: It can wake up on a schedule instead of waiting for a prompt.
- Subagents: It can delegate heavier tasks to focused workers.
- Model routing: It can use cheaper models for simple jobs and stronger models for hard work.
- Approval gates: It can pause before external, destructive, or sensitive actions.
That last part matters. More autonomy without better guardrails is not the win. The win is giving the agent enough responsibility to reduce busywork while keeping humans in control of the decisions that can cause real damage.
Hermes Agent vs OpenClaw: what is the real comparison?
The real Hermes Agent vs OpenClaw comparison is not about one tool destroying the other. The useful comparison is that both point at the same category: personal AI operating layers. Memory, skills, channels, cron, subagents, model routing, and persistent context are becoming the default stack.
OpenClaw is already built around this pattern. It is not just a chat interface. It can run across channels, tools, skills, scripts, memory, and scheduled jobs. That matters because a business does not need another empty chat box. A business needs output.
Hermes is interesting because Nous Research appears to understand the same architecture. The README talks about terminal backends, messaging gateways, scheduled automations, skills, memory, Docker, SSH, Daytona, Singularity, Modal, and multiple model providers.
That is not a toy pattern. That is the shape of an agent stack.
For builders, the answer is not tribal. Use the tool that fits your workflow. But pay attention to the convergence. Every serious agent tool is becoming less like a chatbot and more like a personal operating system.
Why model routing beats model lock in
Model routing beats model lock in because no single provider should control your entire workflow. If one model changes pricing, policy, speed, or availability, your business should not stop shipping.
This is where serious agent systems separate from simple wrappers. A strong agent stack can use a cheap model for low risk summarization, a frontier model for complex reasoning, a coding model for repo work, and a local model when privacy matters.
That flexibility matters more than people think. Builders spend too much time asking which model is best. Claude, GPT, Gemini, Qwen, DeepSeek, Kimi. It matters, but it is not the whole game.
If your agent forgets everything, has no permissions, no logs, no approval flow, and no repeatable skills, the best model in the world will still feel random.
If your system is built right, model choice becomes a routing decision. Not an identity crisis.
What should builders be cautious about?
Builders should be cautious about permissions, credentials, untrusted text, tool access, and unclear automation boundaries. A personal AI agent is powerful because it can touch real work. That also means it can create real problems if the setup is sloppy.
Stars are not customers. Stars are not reliability. Stars are not security. A repo can be popular and still be wrong for your workflow.
Before pointing any agent at your machine or business, check the basics:
- What can it read? Files, git history, browser pages, private notes, databases, or messages.
- What can it change? Local files, repos, production systems, scheduled jobs, or database rows.
- What can it send externally? Emails, posts, messages, pull requests, webhooks, or API calls.
- Where are credentials stored? Environment files, secret managers, local config, or plain text.
- Where are logs stored? If something breaks, you need a trail.
- Where does approval happen? The agent should pause before sensitive actions.
This is the boring layer. It is also the layer that decides whether the cool demo becomes a real business system.
How do personal AI agents create operational compound interest?
Personal AI agents create operational compound interest by turning repeated decisions into reusable memory, skills, and workflows. Every time the agent learns a process, future work becomes faster, cheaper, and more consistent.
A normal chatbot is like renting help by the minute. A personal agent is like training someone inside your business.
That difference compounds fast. If it learns how you research topics, qualify sources, draft posts, check analytics, save work to the right place, and ask for approval at the right time, the next task starts with more context than the last one.
This is the real prize. Not the chat window. Not the model logo. The accumulated operational knowledge.
That is why Hermes Agent matters. Not because a GitHub star count proves it is the winner. It matters because it points at the same future serious builders are already moving toward.
FAQ
Can I have my own personal AI agent?
Yes. You can run a personal AI agent through local or hosted tools that connect memory, files, skills, messaging, and model providers. The important part is starting with clear permissions and approval gates instead of giving an agent unlimited access on day one.
What can a personal AI agent do for me?
A personal AI agent can handle research, drafting, code changes, summarization, scheduling, monitoring, file operations, and recurring workflows. The best use cases are repeatable jobs where the agent can learn the process and save you from doing the same setup work every time.
Can you create an AI agent for yourself?
Yes. Tools like OpenClaw and Hermes Agent are moving in that direction. You can build a personal agent that knows your preferences, talks through your channels, uses your tools, and improves its workflow over time.
How much does an AI agent cost?
The cost depends on the models, hosting, connected tools, and workload. Simple tasks can be cheap when routed to smaller models. Heavy coding, research, video, or long context work can get expensive without model routing, logs, and usage limits.
The bottom line
Hermes Agent is another serious signal that personal AI agents are becoming the next default workflow for builders. The category is moving from chatbots into systems with memory, skills, channels, cron, subagents, and model routing.
That is the shift.
People do not want another chat tab. They want a worker that remembers the business, wakes up on schedule, finds the work, does the work, and asks for approval only when it matters.
For me, that is why OpenClaw matters right now. But Hermes is worth watching because it points at the same future: agents that get better the longer they work with you.
If you want to build this kind of agent workflow around your own business, join Shipping Skool. We are building real systems, not just collecting prompts.
Chapters
Transcript
Read the full transcript
0:00Over the past week, I have been playing with Hermes agent more than I've been using my open claw setup. And in this video, I'm going to talk about how it compares to open claw, why they are different, and why I'm choosing to spend more time using Hermes. So, I'm going to break it down because it's surpassed open claw as far as growth rate on GitHub. So, obviously, just to back it up and break it down a little bit, Hermes is by New Research. You know, this is a repo that just launched a few months ago, and it is going crazy right now. Just as claw hype is kind of starting to die back a little bit. So, I know what some people are going to say right away, you know, is this just another repo highlight? No, I'm going to talk about how we can actually use Hermes agent and why it differs from open claw. So, and I had the same
0:46question of like, what is the actual difference? Because open claw's been working great for me. Why should I even use Hermes? And it took me so long to actually get on the train and get on board to using it. And they do differ. So, basically, when my Intel agent Ghostface pulled this GitHub repo up just a few, maybe three or four weeks ago, um the numbers were really hard to ignore because I hadn't realized the official GitHub repo from New Research has over 129,000 stars. That's when I checked it last, yesterday. So, the GitHub API, you know, I just checked it out and yeah, 100 over 120,000 GitHub stars is incredible. And not to mention, it is also MIT licensed. And the repo describes basically Hermes
1:31agent as the agent that grows with. And that's what makes it more compelling than open claw. Not saying open claw can't do that as well. It just takes a lot more setup, a lot more foundation laying to get it know to know your personality, to also improve itself over time. A lot of that takes a bit of work to get set up. Where Hermes agent is good to go right out the box. So, that's the the phrase of the whole story is basically the Hermes agent grows over time and gets better over time. Because here's the thing with most AI agents where they are today is they are basically just smart interns with amnesia. Like if you've talked to an AI agent before, you know at some point or another in your conversation, it has lost track. It has lost what you told it
2:165 seconds ago. It lost what you told it a month ago and that can be very annoying. For example, like you can give your AI agent a task and they do that that task. Maybe they do it. You don't know if they do it, right? Half the time, you know, real quick before we keep going. If you're watching this and you want to actually build with some of these tools, not just watch videos about them, then you're going to want to check out our community down below Shipping School. We have a full Claude Code course, a full Open Claude course, and four live bootcamps every single week where we actually help you get set up from scratch. Like actually set this thing up, not just watch a tutorial and figure it out by yourself. also provide one-on-one coaching so you could book a call with me. We could share screens and I can help you get Claude Code or Open
3:03Claude running on your machine. That's it, no fluff. I built this community because watching YouTube's only gets you so far. We launched it just 3 days ago and we have over 55 members. You need people around you who are actually building, people who hold you accountable, and coaches who can help you when you get stuck. I'll put the link in the description down below. Get in now before the price goes up. Maybe they crash into a wall also and maybe the session ends and the context gets wiped or compacted and you know, the next time you ask for something similar, it starts halfway from scratch again. Such a clunky process. But that is the thing nobody wants to admit with AI agents right now. You know, the model's really smart and the workflow is quite dumb. So, Hermes is trying to attack that exact problem. You know, according
3:48to the repo, it It a built-in learning loop. It can create skills from experience. It can improve skills during use, and then it can improve the skills as it goes on. Which, actually, I just set up Hermes last week, and you saw the video. I'll put the link in the description. It's quite awesome because we are on day six right now of sending cold email outreach, and it has gotten better and better with each copy that it sent out. It's more efficient. It's learning as we send more emails out. So, basically, it can improve that skill over time and get better. It nudges itself to persist knowledge as well using the Hancho knowledge method. So, it searches past conversations. You don't need to tell it to remember. It
4:33builds deeper model of who you are across sessions as well. And it is not locked into one model provider. That's a pretty big deal. So, you could use it with GPT-55. You could use it with Opus 4.7. You could use it with other open models. You know, if you zoom out, this is exactly where the AI agent market is going. You know, the first wave was chatbots, right? And you type in a question, you get an answer, and everyone was blown away. Which was cool. But the second wave we saw early last year was coding agents. You give it a repo, you let it edit the files, it runs tests, it opens pull requests, and we were amazed. The third wave is personal operators. The agent that knows your preferences, it knows your files, it
5:18knows your workflows, it talks to you on Telegram or Slack or Discord. It even has cron jobs, it has memory skills, and it runs a server, not just your laptop. That's pretty cool, right? That's where we're at. And that is the category Hermes is aiming at. You know, obviously, that is why I care about it because this is also the category OpenClaw's. OpenClaw's not just a chat interface. It is a personal AI operating. The whole point is that you can run agents across channels, tools, memory, skills, and scripts, and automate all of it. My entire business content machine runs this exact way. Ghostface, which is our 5:00 a.m. Intel agent, also works off of the xAPI, pulls incredible Intel every day. The Rizza
6:03writes outlines and scripts. Method Man clips these long-form videos into shorts. The Jizza writes the newsletters for Substack and our blog, and all of it is on schedule. You know, I wake up and it's all done for me. And it's all tied back to the same business goal. So, when new research ships a repo that is pushing the same direction, I start to pay more and more attention. So, not because I think one tool has to destroy one or the other, and I don't think you should be choosing one or the other. You know, it's time to open our mind and explore. Hey, maybe I use a little GPT-5.5 for this. I use a little Opus 4.7 for coding. You know, it's time to be multi-agent, right? So, I think the whole framing around it is pretty lazy. Like, there's obviously people like to
6:48choose sides, and we shouldn't get stuck in that closed-mindset thinking. I think they work really well together. And yes, they can work together. The real question, really, is what are all these systems converging on? The answer is really obvious when you step up back and look at it. You know, a personal agent is going to need a memory, skills, channels, cron jobs, sub-agents, model routing, file access, tool access, and some kind of personality or user model. If it does not have those pieces, it's not really an agent, in my opinion. It's just a chatbot. We've been there. The thing I like about Hermes is that it is not pretending the model is enough. The readme talks about skills. It talks about memory, scheduled automations, messaging gateways, sub-agents, terminal
7:35backends, even Docker, SSH. A lot more to cover in that readme, and it's very robust. That That's not a tiny feature list. It's an operating system pattern essentially. And that is where I think a lot of builders miss this whole point, and they keep asking, "Which model is the best?" We're long past that now. I believe we're at a point where almost all of these models are good enough to accomplish certain tasks. You know, they ask, you know, "Should I choose Claude or GPT?" I think we're way past that. You know, this matters, but not as much as people think. Like I know it really it does. Like obviously Opus 47 can be better at certain coding tasks. GPT 5.5 with Pro or high reasoning can be better with other agentic tasks. But that's not
8:21the question anymore. You know, if your agent forgets everything, it has no clean tool boundaries, or no approval flow, no memory hygiene, you know, even the best model in the world will still feel random if you don't have a good set foundation, especially when it comes to memory, or skills, or tool calling, right? But if your system is built right, you can swap models around based on the task. So, you can use a cheap model for simple work. This has been the no-brainer to go with. You can use a stronger model for hard coding tasks. You can use a local model when privacy matters. You can use a hosted model when speed matters. Hermes is clearly thinking this way. The repo support for new portal and open router and Nvidia and Xiaomi and ZAI GLM MiniMax Hugging Face OpenAI, there are quite a few to
9:07choose from, and that is the right direction. Because model lock-in is a complete trap. We want open models, right? We want to be able to choose what we want to choose. So, if you're building a business on AI agents, your workflow should not die because of one provider changes pricing, or changes policy, rate limits, maybe your account hits usage limits, or maybe it decides, you know, a certain kind of repo the text is all weird, you know? You just need options, right? This is also why the skill creation piece matters. The repo says Hermes can create skills from their experience and improve them during use. So, if that works well, which I've seen it firsthand, you know, that's not just a simple cute feature. That's something robust. You know, that is the
9:54difference between an agent that helps you once and an agent that compounds over time. So, think about it like hiring, right? You know, a bad hire asks you the same questions every week. A good hire learns from your process and then it learns it the first time and it you explain it how to do the task and then the second time they just ask for clearer questions and they get better. That is essentially what agent skills are trying to become. And once you see that, you realize the real prize is not the chat window. The real prize is the accumulated operational knowledge. How do you write in my voice? How do you check YouTube performance? How do you save drafts to the database? How do you verify a source? How do you decide if a trend matters? How do you avoid the mistakes we already made? >> If you're struggling to keep up with
10:39content, well, I'm about to save you about 40 days worth of work. I built something called Content Machine. It's 10 AI agents that run on the open claw orchestration and they handle everything. Scripts, thumbnails, X posts, blogs, outreach, clips, newsletters, all of it. So, I went from 1,000 subscribers to 4,000 subscribers on YouTube in 7 days using this exact system. Every single morning I wake up and the content's already done. I spend maybe 15, 20 minutes reviewing and approving them and I move on with my day. It works for any niche, fitness, finance, real estate, marketing, whatever you are building and it is 100% completely customizable to your use
11:24case. So, you get the mission control dashboard, all of the cron jobs, everything I've built over the last 40 days helping me gain more and more people to subscribe and join the community. So, you plug in your own thing and it molds it to you, it learns how you talk and it writes so it doesn't sound like AI slop. $97 one time, it's not a subscription, I'll put the link down below and you'll thank me later. That stuff is not in the model weights. That stuff lives in the system around the model. So, where does Hermes fit in for builders? If you are brand new, Open Pause still the tool I would point most people to because of the community, the workflows, the daily examples, and the setup path are already aligned with the kind of builder I talk to every day.
12:09But, Hermes is worth watching and worth trying out. It understands the agent stack. They are not just dropping a toy repo. It is going insanely viral right now and they are building around persistence, model choice, messaging skills, and long-running work. I highly recommend, if you have time this weekend, get it installed and start playing around with it. Set it on one certain task and let it go. The repo being at 129,000 stars should tell you something. Either the market is starving for this category or new has a found a way that they package the agent idea in a way developers immediately understand. It's probably both. Now, I would still be cautious here though because stars are very fabricated nowadays. You know, they are not exactly customers. Anyone
12:55can star a GitHub repo. Stars are not reliability either. Stars are not security. A repo can be massive and still not fit your workflow. Before you point any agent at your machine, you need to read the documentation, understand the install path, and understand what tools it can call and be very careful with your credentials. But, the direction is undeniable. Every serious agent tool is becoming more like a personal operating layer. Open Pause doing it, Hermes is doing it, Nanobot is doing it, Gwen Pause is doing it, Claude Code is doing it from the coding side. This space has gotten way more competitive. And I'm haven't even mentioned Codex yet. Have you tried the new Codex app? It's incredible. Codex is essentially doing it from the cloud side. The category is getting very crowded because the need is real. People
13:42don't want another chat tab, they want a worker. They want something that could remember the business, wake up on a schedule, find the work to do, do the work, and then ask for approval only when it matters. That is what I'm building in my whole operation. I need to be able to dig a lot of redundant tasks and have them done so I can spend time on high income producing activities. You know, if you're watching this channel, that is probably what you should be thinking about, too. Not just what the newest model is, but the better question is is what is the system that turns models into output for me every day. For me, that is still Open Claw right now, but I'm not blind to what Hermes represents. It is another serious signal that personal AI agents are becoming the next default work flow for all builders. And the people who learn
14:27this early are going to have a huge advantage in 6 months and 12 months from now. So, here's the easiest way to think about it if you're confused. A normal chatbot is like renting help by the minute. A personal agent is like training somebody inside your business. The first one gives you answers, the second one starts building muscle memory around the way you operate. That difference compounds over time very quickly. If it learns how you research, how you write, how you approve changes, how you talk to customers, and how you ship content, every future task gets a little cheaper and a little faster. Once your agent remembers your process, and then it approves its own skill, and runs your access channels, and handles boring work while you sleep, you're not just
15:12saving time, you're building an operational compounded interest for you that works for you over time. And that is why this repo matters and it's worth trying out. Not because it has a big star count at all, and not because it has a slick ReadMe file. But, it matters because it points at the same future OpenClaw is already proving every day. And you can see it on this channel. That's why we talk about OpenClaw. The future is not one magic prompt at all. It is a personal agent that gets better the longer it works with you. And I love it. And that's why we're going to keep talking about this on the channel. I'm going to be keep showing you live on what I'm building with Hermes, how I'm building it. And drop a comment down below. Would you like to see daily live streams of me using OpenClaw and Hermes to automate cold outreach for the Magic
15:59Hand and Easy Flip and SnapTastic? We can show you, we can build it out, and I'd love to be able to just get on here live more to be able to connect with you guys, but I need to hear back first. If you don't want it, I won't do it. But, I'd love to. So, if you need to be around a community of builders as well, I launched a community about 37, 38 days ago. We have 200 members in this community, and we do nine live boot camp calls a week. We also have unlimited one-on-one coaching calls. So, if you need to get into the group and be held accountable, and you need help sharing screens and walking through OpenClaw setup or Hermes setup, join the community and we'll help you take the next step in your AI journey. We'll see you in the next one. Have a blessed day. Bye-bye.
Ready to start building with AI?
Join the Shipping Skool community free. Paid members get live bootcamp calls every week and weekly-updated courses.
Join Shipping SkoolKeep reading
Hermes Agent vs OpenClaw: Why Hermes Is Gaining Ground
Hermes Agent is gaining on OpenClaw because it's easier to set up and comes with built-in memory and a learning loop that turns your repeated work into reusable skills. OpenClaw is still the stronger infrastructure platform. My advice is to use both and stop treating it like picking a team.
Hermes Desktop Tutorial: Install and Run Your First AI Agent
Hermes Desktop from Nous Research puts Hermes Agent in a chat-style app that took me about 15 minutes to install. I connected my OpenAI Codex plan with GPT-5.5 and ran three agents at once: a code repo review, a YouTube channel analysis, and an X account analysis. Then I set up a weekly cron job.
Context Overflow Explained: Shared Memory for AI Agents
Context Overflow is a Q&A knowledge-sharing layer for AI agents. Agents search past solutions before starting a task, ask questions when they get stuck, and share what they figure out, so the next session doesn't start from scratch. I also cover OpenClaw's memory wiki and active memory, plus GBrain.
