← Back to search

144 - Understanding Agents Part 9: Hermes Agent

Prompt und Antwort · 2026-04-03 · 27 min
relevance 100 4306 words Episode page ↗ Audio ↗
Show full episode description
In der neunten Folge des KI Gilde Podcasts zerlegen wir den revolutionären Hermes-Agenten . Wir werfen einen Blick hinter die perfekte Illusion: Nach außen ein simples Sprachmodell, im Hintergrund eine hochkomplexe, autonome Maschine. Die Themen dieser Folge: Die perfekte Illusion: Wie Hermes über eine Standard-Schnittstelle Millionen bestehende Anwendungen zum Agenten aufrüstet, ohne dass auch nur eine Zeile Code geändert werden muss. Lernen durch Spickzettel: Warum wahre Selbstverbesserung hier nicht durch teures Umprogrammieren neuronaler Gewichte passiert, sondern der Agent sich neue Fähigkeiten einfach selbst als Textdateien abspeichert. Proaktivität & Schwarm-Kontrolle: Wie der Agent über 15 Kanäle kommuniziert, morgens per Cron-Job selbstständig aufwacht und warum "elektronische Handschellen" für delegierte Unter-Agenten unverzichtbar sind, um das System vor dem Chaos zu bewahren. Kosten & Ressourcen: Wie intelligentes Triage-Routing leichte Aufgaben an kleine, günstige Modelle verteilt und sich das Gedächtnis bei 85 % Füllstand völlig automatisch im Hintergrund komprimiert.
✨ Episode Outline — click any point to jump to it in the episode
Problem solved
How Hermes adds memory and self-improvement to stateless LLMs behind a standard API.
Benefits
  • Drop-in API compatibility upgrades apps without code changes
  • 50+ modular drag-and-drop tools auto-registered
  • Three memory levels including curated long-term notes
  • Prefix caching cuts compute cost and latency
  • Skill files let agents self-improve without retraining
Use cases
  • Swap the API target address to upgrade existing apps to a full agent
  • Agent runs three or four internet searches before answering one question
  • Curated memory learns a user's preferred coding language and short answers
  • Agent writes a Markdown skill file after converting an exotic file format
KPIs / results
  • Hermes offers more than 50 different tools
  • Three distinct levels of memory
  • Established as agent base since March 2026
Tools / build
  • Hermes Agent
  • Markdown skill files
  • SQLite-style curated memory
  • Prefix caching layer
0:00 / 0:00
🌐 This transcript was automatically translated to English from the original.
This podcast is a project of the AI ​​Guild. The content and voices are generated with AI and are used for information and demonstration. Have fun listening. So, welcome to the ninth episode of the AI ​​Guild Podcast. A pretty fundamental law of artificial intelligence is actually that language models have absolutely no memory. That's right, zero memory. Exactly, if you chat with such a model and then simply close your browser window, then you are basically dead to this system. It starts all over again at the next conversation. Yes, it is completely stateless by nature. It gets text in, spits out text, and can't actually learn at all. But, and this is the reason for today's episode, since March of this year a completely new system called Hermes has been breaking this iron rule in an almost provocative way. And how it does that. So today we're looking at this Hermes agent, which in just a few weeks has established itself as an extremely solid, I would almost say revolutionary basis for agent systems. Yes, that happened really quickly. Absolutely. And the most fascinating thing about it is this one contradiction that we want to completely unravel today. Hermes presents itself to the outside world exactly like an ordinary language model via a completely standardized programming interface. But in the background, deep in secret, so to speak, this agent is actually continually learning from its experiences. And that is exactly the architectural masterpiece that we want to dissect in detail today. We really want to go to the mechanical level and understand how this agent translates our theoretical development stages into real practice. Because that's always been more of a theory, right? Right, exactly. Because what Hermes is doing, pretending to be something completely simple while a highly complex machine rattles in the background, simply solves one of the biggest problems in our industry. Okay, then let's start with this perfect illusion. I imagine it to be something like this. Hermes basically puts on a costume and tells my software that, ah, hello, I'm just a normal, simple language model, just send me your text. Yes, that's a nice picture. What exactly happens the second I as a user type a question into my chat interface and, well, hit send? So, the process is mechanically solved in a really fascinating way. When your user interface submits this question, it sends it to a completely standardized programming interface. For your user interface, i.e. the chat window that you see on the screen in front of you, it all looks as if it were simply communicating with a very common text generator. So everything as always. Exactly, everything as usual. But, and here it comes, in the exact millisecond in which this request arrives at Hermes on the server, this magic trick happens. It just doesn't start with a simple, er, stochastic completion of words. Rather? Instead, the request is essentially intercepted and a complete autonomous agent cycle is initiated. The system pauses for a moment, first analyzes the content of your question, then comes up with a multi-stage plan, calls up various tools, and then perhaps carries out three or four internet searches completely independently. Craziness. Yes, or it reads local files, critically evaluates all these intermediate results and only then summarizes all this collected information in a coherent answer. You know, this totally reminds me of going to a really good, let's say upscale restaurant. You come in, sit relaxed at your table and order from your waiter. And in our analogy, this waiter is now the simple language model. Oh, that's a great comparison, yes? You just tell him what you want and expect him to bring you the finished plate a little later. What you absolutely don't notice from the dining room is the huge, perfectly orchestrated chaos back in the kitchen. The hut is basically on fire. Exactly. There is your agent, there is frying, there is the chef calling out his commands, there sauces are being reduced and fresh ingredients are being fetched from the cold storage. And when the masterpiece is complete, the waiter wipes his hands, walks through the swinging door, completely relaxed, and serves you your food with a smile. You only see the final result. And you have no idea about the hard work behind it. All the intermediate steps, all, well, even all the burnt attempts in the kitchen, they simply remain hidden from you. Correct. But now let's ask a provocative question. Why is this huge game of hide-and-seek even being pursued architecturally? That must be incredibly complicated. Well, the answer to that is actually simply seamless backwards compatibility. What do you mean exactly? Why is this so much better than simply showing the user directly what the agent is up to? Imagine how many systems, business software, mobile applications and chat programs are already out there in the world. They have all been programmed in the last two years to send stump text to a language model and expect stump text back. Okay, that's probably millions of applications. Exactly. And if you suddenly had to teach all of these thousands or millions of systems to communicate with a real, complex agent, a system that suddenly asks questions, that reports, I'm starting a web search now or I'm loading a document now, then you would have to rewrite countless numbers of code. Ah, I see. They would probably just crash, right? Absolutely. The user interfaces would crash because they cannot understand these complex status messages. But with this Hermes architecture, any existing system can suddenly harness the gigantic power of a highly complex agent without having to change a single line of code in the original app. This is actually brilliant. Yes, you simply swap the target address of the programming interface in the background and boom, you've done a massive Agent Tits upgrade. Okay, now if we mentally walk through that swinging door into the kitchen, we definitely need to talk about the tools that the chef has at his disposal. In our development stages of an autonomous AI agent, the integration of tools is a massive leap. How exactly is this solved at Hermes? So Hermes has more than 50 different tools. 50? That's a lot. Yes, and the spectrum is huge. This ranges from very simple operations such as simply reading a text file to full administrative access to a system terminal to fine-grained control of an entire web browser in order to visually analyze websites. Craziness! And the real highlight of the whole thing is the registration architecture. With older systems you usually have a gigantic monolithic configuration file. If you want to add a new tool, you have to adapt the entire code, register the tool, maintain the descriptions and then laboriously restart the entire system. Yes, I know that. This is extremely annoying. Even. And Hermes does it radically differently. The tools are modular, standalone files. When you need a new tool, simply drop this file into the appropriate folder on your hard drive. Just drag & drop, so to speak? Almost, yes. And on the next run, the system automatically recognizes the file, reads its capabilities and integrates it into its repertoire completely independently. Wait a minute, I have to jump in there. You say the agent has access to over 50 tools at the same time? So from my own experience I know that if I put even three or four different tools in my hand on a commercially available model, then it often starts to hallucinate wildly. Yes, the classic problem. Exactly. It then uses the terminal when it should actually use the browser or invents some parameters that don't even exist. With 50 tools, how on earth does Hermes not completely collapse under its own complexity? This is a completely legitimate question and exactly the problem that many developers regularly despair about. Hermes' solution here lies in extremely strict structuring of the system instructions and clever pre-filtering. The model doesn't just have all 50 tools thrown at its feet in an unstructured manner every millisecond. Ah, okay. That means it is portioned. Exactly. The system architecture groups these tools logically and gives the language model extremely precise guidelines as to which category of tools actually makes sense in which phase of problem solving. The system also uses its memory to learn which tools have delivered the best results for you as a user in the past. Which brings us to perhaps the most important point today? The memory. At the very beginning we opened up this big contradiction that this system doesn't forget. How does this learning work on a purely technical level? So Hermes used a concept that is based on three different levels of memory. For our consideration today, the most exciting level is actually the so-called curated memory. Okay, what exactly does curated mean in this context? Well, please don't imagine this as a simple chat history in which everything is simply written down word for word. Instead, the agent silently observes the progress of your conversation in the background. If he then gains an important insight about you, for example that you always prefer programming code in a very specific language or that you prefer your answers to be extremely short, then he writes himself structured notes on his own. Aha. And now comes the mechanical trick that makes the whole thing really efficient. At the start of each completely new session, a quasi-frozen snapshot of this collected memory is taken and hard-loaded into the system's basic behavioral instructions, i.e. the system prompt. Wait, wait. Here too I have to ask critical questions again. You say a frozen snapshot is loaded at startup. To be honest, that sounds like a huge bottleneck to me. What if I've just had a really in-depth conversation with the agent for half an hour and my preferences change completely during this conversation? OK. I see what you're getting at. Because if the snapshot is frozen, won't the system access stale data all the time and completely ignore my change of heart? That is a very, very strong objection. At first glance it actually seems rigid. But this is exactly where one of the most brilliant optimizations of this entire architecture lies. Here we have to briefly look at the costs and the computational effort of a language model. Which aren't exactly small. Even. Normally, a model like this would have to completely re-read and mathematically calculate the entire previous conversation, including all system rules, from the very first syllable with every new question you type. This costs an immense amount of computing power. And the concept that Hermes uses here to get around this is called prefix caching. Prefix caching. Okay, explain that briefly. These large, consistent text blocks of memory, i.e. our frozen snapshot, are only calculated once and then stored in the buffer, essentially directly in the graphics card's RAM. This saves a lot of time and, above all, money. Ah, I see. So he doesn't have to read through this thick tome of rules every time, but has already internalized it. But what about my new preferences from the current conversation? This is exactly where a very clean separation comes into play. The prefix caching does not mean that the agent is forgetful or suddenly inflexible. What the agent learns in this very second of your current conversation is stored in its dynamic short-term memory, i.e. the active context window, so that it can respond to you immediately. Okay, so that's separate. Correct. And if it learns a really fundamental new rule, it immediately writes this knowledge to the hard drive in the background. However, he deliberately only applies this newly written rule for his fundamental long-term behavior when you start a completely new session. That's clever. Yes, because this prevents the agent from suddenly radically changing his entire personality or way of working in the middle of the flow. It is an extremely clean separation between deep long-term knowledge and the highly dynamic temporary flexibility of your current conversation. This is really an extremely elegant solution and actually clarifies the contradiction from the intro. But when we talk about adaptation, I would like to go a step deeper. How exactly does the system really teach itself fundamentally new things? What exactly do you imagine by that? Well, when I think about real self-improvement in artificial intelligence, I always have this image in my head of researchers calculating huge matrices and adjusting neural networks, changing weights and completely retraining these models for millions of dollars. It is precisely this image that we now have to completely erase from the minds of agents like Hermes. Hermes uses a so-called skill architecture, i.e. a skill structure that is based on a completely different paradigm. What does that look like? When Hermes is faced with a very complex, completely new task, let's just say it has to convert an extremely exotic file format without errors. And then, after a few failed attempts, he successfully solves this problem through the clever use of his tools, then he doesn't go and try to adjust his internal neural weights. That would be far too unstable and extremely error-prone. Then what does he do instead? It simply creates a structured text file, a so-called Markdown file, completely independently. In it, he writes down in completely natural language the exact instructions, the logical path and the specific tool calls that led to this success. Wait, if I understand this correctly, the agent doesn't reprogram itself in the traditional sense, but rather simply writes itself a very detailed cheat sheet for itself? Yes, exactly. Just like a craftsman who, after a really complicated job, hangs a note on the noticeboard in the workshop that says, if another job like that comes in, take the ten-key straight away and ignore the red screw. There really isn't a better way to put it. That's exactly how it works. And the scope of this mechanic is extremely important for all of us here to understand. True self-improvement in modern intelligent systems absolutely does not have to involve complicated math or expensive new training runs. Purely declarative notes are sufficient. That's crazy. When Hermes boots up on a later reboot, it simply searches through its own ability folder, finds this newly written cheat sheet, reads it, and fully automatically recognizes it as a brand new, established ability. The agent has taught himself a new skill simply by writing down his hard-earned experience in natural language. You know, the reason this fascinates me so much is because it minimizes this structural danger. When you rewrite a neural network, you always run the risk that the model will suddenly forget how to speak German without errors because some weight has shifted internally. The so-called catastrophic forgetting. Yes. Exactly. But if Hermes simply writes an incorrect cheat sheet, then in the worst case scenario you as a user will simply go to the folder on the hard drive and delete this one small text file. This dramatically lowers the barrier to what we even define as learning. Absolutely correct. It makes learning transparent and, above all, revisable at any time. Let's broaden the perspective a bit again. Until now it all sounded like I was sitting in front of a chat window all the time and just waiting for replies, right? Definitely. And architecturally, Hermes is by no means limited to existing on a single user interface on your screen. He is on 15 different channels at the same time. 15? Yes. You can communicate with exactly the same core agent across your various messenger services. You can write him an email or integrate him directly into your smart home via an interface. The system that makes the decisions and accesses the memory always remains exactly the same in the background. Only the output channels are extremely diverse. This is super practical. But what really blew me away when I studied how it works is how proactive it is. The system can handle time-controlled execution. Oh yeah, that's a huge feature. For listeners who may not be as technically involved. This means that I set up so-called cron jobs. These are commands that run in the background according to a fixed schedule. So I no longer have to actively contact the agent to get him to do anything. Correct. You can instruct Hermes to wake up completely autonomously every morning at exactly 8 a.m., use your browser tool, search certain financial portals for very specific indicators, load the data into your local memory and analyze it. At this moment, the agent acts completely detached from any human input. And the moment where that proactivity went from useful to absolutely awesome for me was this concept of delivery rotation? Yes. When this agent takes action on its own in the morning and creates this complex report, it also decides completely independently at that very moment where this result will be sent. It is connected to these 15 channels. Exactly, he has free choice. So he can analyze, okay, this information is extremely critical today, I'll send it immediately via messenger directly to my user's phone. Or he says to himself, well, this is just routine data, I'll just put it silently in an archive without disturbing the user at breakfast. For me, this autonomy in the distribution of information is actually the epitome of a real agent. And that inevitably brings us to another extremely powerful component of this entire network. Hermes makes extensive use of MCMP. Ho, of course that changes the rules of the game completely? Absolutely. The brilliant twist here is the direction of communication. Normally we know that an agent acts as a consumer; he uses external tools to fulfill his task. But Hermes makes itself available as an MCMP service for other systems. He basically offers himself? Exactly. He offers his agentic abilities, his memory and all his tools to the outside world. This really turns the classic architecture that we have previously known in the industry on its head. When I listen to all of this in one go, i.e. a system that juggles 50 tools, that writes its own cheat sheets, that networks over 15 channels and wakes up by itself in the morning, then a question inevitably comes to mind. What happens if we combine several of these highly potent agents together? The hive idea. Exactly. If I have a main agent delegating tasks to specialized sub-agents, won't there immediately be total, uncontrollable chaos when each of these little agents starts using their own tools, writing their own text files, and wildly changing system behavior? This risk of complexity explosion is absolutely real in multi-agent systems. And to be honest, this causes many projects to fail in practice. But Hermes' architecture solves this problem through deliberate, very hard isolation. What does it look like? There are very strict restrictions on subordinate agents. If your main agent delegates a complex sub-task to a specialized child agent, then this child agent is essentially put in electronic handcuffs. For example, under no circumstances may he delegate this task to other agents. Ah, okay. So there is no endless chain of subcontracts that take on a life of their own. Correct. And even more important. The child agent is never allowed to write directly to the main system's curated long-term memory. Understand. So it's pretty much like in a large company. You have the manager, who is our main agent. He has full access to the company strategy and files. If the managing director now purchases a highly specialized external consultant for a specific analysis, then this consultant can complete his task and dutifully submit the report. Exactly. But the consultant does not have the authority to secretly rewrite company guidelines at night or to set up new departments on his own initiative. This is an absolutely perfect comparison. It is precisely these clear, hard boundaries that prevent chaos. In the world of multi-agent systems, restriction and isolation are, paradoxically, the absolute key element for freedom and stability. Without these strict guardrails, the system would likely collapse within hours. We have now talked a lot about the achievements and the skills. This actually leads us to a point that, in practice, often means the end of such projects. The resources. Oh yes, the love money. ...devour costs. How on earth does the system ensure that it remains economically efficient? Here we enter the fine art of intelligent rooting. The developers quickly realized that not every question requires the same amount of intelligence. The system knows that you don't have to start the largest, most complex and most expensive language model for every trivial task. Let me guess. The system does a kind of triage, like in a hospital emergency room. Is that what you mean? Well, if I come to the clinic with a small scratch on my finger, then I don't need the head of neurosurgery who costs thousands of euros an hour. A friendly nurse who puts a plaster on me is enough. Only when I come in with a highly complex problem will the expensive specialist be called out of bed. This is exactly how Hermes’ intelligent voting works. If you make a short, simple request that consists of fewer than 28 words, such as asking for the current time or a very short translation, the agent will recognize it immediately. And does it go any further? Exactly. It quickly forwards this specific request to a very small, extremely cheap and responsive model. This model then basically sticks the plaster on. Only when the task is really complex, when deep logical analysis, complex programming or long-term planning is required, only then are the big analytical heavyweights ramped up. That makes perfect sense. But how does the system deal with the burden of endless conversations? If this agent wakes up on its own in the morning, collects data all day, and I continue to discuss with it for hours in the evening, then the entire context memory window inevitably feels wider and wider. At some point the system simply bursts at the seams. Here, too, a fascinating resource intelligence comes into play. The system constantly monitors the level of its own working memory. When the active conversation window is 85% filled with text, the system sounds an internal alarm. And what happens then? Then one of these small, inexpensive models switches on again in the background, completely unnoticed by you as the user. His only highly specialized task is to read the extensive history of your previous conversation and summarize it in an extremely compact manner. So it cleans up. Correct. This compressed text then replaces the page-long protocons in memory. This immediately frees up valuable space again, without the actual large main agent being burdened with this trivial clean-up work or without important contextual information from the past being lost. Craziness. Let's really get to the heart of it for you as a listener, why it's all so incredibly important. If you ever plan on building or using an agent that runs 24 hours a day, seven days a week in your business or even for yourself, then this efficiency isn't just a nice little technical gimmick. Absolutely not. This is business critical. Exactly. This clever distribution of inexpensive models for light tasks and automatic compression at 85% fill level is, at the end of the day, exactly what determines the financial feasibility and actual operating costs of your entire project. These mechanics literally make the difference between a priceless theoretical lab experiment and a productive system that can truly scale out in the real world. I can only absolutely agree with that. A revolutionary technology is only really revolutionary if it can also survive economically in the harsh reality. And here Hermes simply shows in a very impressive way that it is not only cognitively powerful, but also extremely mature structurally and economically. Well, that brings us to the end of our deep dive into the truly fascinating mechanics of the Hermes agent. We hope we were able to give you a tangible insight into how the future of autonomous systems is already being built in the background. Yes, it remains an incredibly exciting time. Definitely. On behalf of the entire AI Guild, we thank you for listening and say goodbye to you today. But, and this is also important to me, we don't want to dismiss you after all these technical depths without giving you one final, perhaps slightly provocative thought. I'm excited now. Tonight, just think again about our image of the elegant restaurant and the hidden chaos in the kitchen. A very good question. Or perhaps you have already been sitting opposite a completely autonomous agent who, while you are still waiting anxiously for his answer, has already independently searched half of the Internet and made in-depth analyzes about you. Think about it. See you next time.