← Back to search

Episode 231 - Hermes Agent

Big Data Hebdo · 2026-07-23 · 42 min
relevance 100 6953 words Episode page ↗ Audio ↗
Show full episode description
Dans cet épisode, Vincent Heuschling et Nicolas Steinmetz décortiquent Hermès Agent, le nouveau venu open source qui bouscule OpenClaw. Au programme : son architecture en cinq couches, sa mémoire persistante, et sa capacité à tourner en local pour préserver votre souveraineté. On parle aussi dans cet épisode du nouveau mode de facturation d'Anthropic, de cas d'usage concrets (daily briefing bot, revue de PR GitHub) et d'un sujet qu'on n'anticipe pas assez : la traçabilité et l'identité des agents qui se chaînent entre eux. ## Chapitres
✨ Episode Outline — click any point to jump to it in the episode
Problem solved
Understanding how Hermes Agent's structured harness rivals OpenClaw and commercial agent runtimes while keeping data sovereignty.
Benefits
  • Ships a pre-built five-layer harness instead of assembling everything yourself
  • Auto-writes skills from observed repetitive prompts and tool interactions
  • Persistent memory with ~10ms retrieval even over huge document volumes
  • Sandboxed permission and execution system limits runaway agent actions
  • Model-agnostic multi-model orchestration, supports self-hosted open models
Use cases
  • Agent reads a list of blogs and delivers a morning news summary autonomously
  • Start a conversation on Telegram/WhatsApp and continue it in the terminal, fully synchronized
  • Use GoCli to list and write Gmail as JSON/markdown without opening broad API access
  • Podcast production: agent connects recent news to past episodes for content ideas
  • Memory system as a RAG backend with ~10ms retrieval latency
KPIs / results
  • Hermes Agent near 100,000 GitHub stars since February 2026 release
  • OpenClaw around 350,000 GitHub stars
  • ~10 milliseconds memory retrieval latency
  • Around 100 bundled skills covering diverse use cases
Tools / build
0:00 / 0:00
📑 Chapters — tap a time to jump there
00:00:00
Introduction et présentation de Nicolas Steinmetz
  • Big Data Hebdo intro: Vincent Heuschling and Nicolas Steinmetz, CTO at Cybeltech
  • Episode topic: hands-on look at Hermes Agent
00:02:00
Hermès Agent, le nouveau rival d'OpenClaw
00:04:00
Harness engineering en cinq couches
  • Five layers: instructions, constraints, feedback, memory, orchestration
  • Hermes observes repetitive prompts and auto-writes skills as editable markdown
  • Arrives with a structured harness instead of building everything from scratch
00:09:00
Mémoire persistante et orchestration multi-modèle
  • Persistent memory switch records all interactions across mediums
  • Start on Telegram/WhatsApp, finish in terminal, fully synchronized
  • Multi-model orchestration, not tied to one provider; ~10ms RAG retrieval
00:14:00
Sandbox, permissions et souveraineté des données
  • Sandboxed permission and execution system prevents runaway actions
  • Nicolas's reluctance: hard to delegate email/Google access to agents
  • GoCli command-line tool handles Gmail as JSON/markdown with scoped permissions
00:20:00
Nouveau modèle de facturation d'Anthropic
00:26:00
Cas d'usage concrets et niveau de contrôle
  • ~100 bundled skills; morning blog-digest agent built very quickly
  • Feature set comparable to AWS Bedrock Agent Core
  • Podcast production idea: agent links recent news to past episodes
🌐 This transcript was automatically translated to English from the original.
Hello everyone, hello everyone, welcome to a new episode of Big Data Hebdo. Big Data Hebdo is the French-speaking data and AI podcast. And in Big Data Hebdo, we try to enlighten you on everything that's moving in terms of technology in data and AI. And so, today, we're going to talk about agents, we're going to talk about AI agents, but we're going to talk about things a little more where we can get our hands on it, since we're going to talk about Hermes Agent. Hermes Agent is the open source agent of the moment. And for this episode, I am accompanied by Nicolas. Hi Nicolas, how are you? Hi Vincent! Well listen, it's going well, hotly, but it's going well. But there, it's okay, there are 6 of us in the morning, so it's okay. Nicolas, you are CTO at Cybeltech, which is therefore a company that is interested in everything that is life science and in supporting culture and everything that follows. Exactly. You presented everything well, nickel. Well yes, I took up what you said last time, you see, I didn't invent anything. Ah yes, you listen, you listen, it’s good. I listen to my interlocutors. And so, for my part, I am Vincent Heuschling, I am the co-founder of this podcast and I am an independent consultant. I support my clients in their Data EIA project to help them imagine, design and also implement solutions that could change their business. And so, let's go. We're going to get to the heart of the matter. In this episode, we finally said to ourselves, we had moved on a little bit, we haven't talked much here in the OpenClaw podcast. And then, there is a new kid who has arrived called Hermès. And Hermès positioned it a little differently. The principle always remains the same as a reminder. OpenClaw is what put Apple out of stock on Mac minis, since the principle was that we were running an agent orchestrator. That is to say, we consumed the resource, for example, anthropogenic to make the agent think. But all the orchestration, all the memorization, the interaction was done on a machine that you hosted at home or also at a cloud host. Moreover, I believe that it was Hostinger that even did a type of deployment for its environment where you could request to have an OpenClaw machine directly installed. And OpenClaw had an interaction mode which was as follows. In particular, you spoke on a Telegram channel with your agent and then your agent did things for you. So, we were in the dialogue of the human with his agent to do things. There have been quite a few things shown on the market. But Hermes Agent and Harry were a little bit different. So already, where OpenClaw was the emanation of a solitary actor, Steinberger, here OpenClaw is an open source lab called Nous Research and which are not necessarily forced to produce scientific papers to be recognized, nor to raise funds. So, in fact, we don't talk about them very much. But they built models. The Hermès 3 range of models which existed in 8 billion parameters up to 405 billion parameters, is indeed one of the super good open source models that you can use. So, and so the thesis that they had is say with an open LLM API and open source tools, you could deploy an agent that rivals commercial offerings. So, it's true that indeed, we could relate this to what Anthropik offers in particular with the ability to have agents who run on their own resources. And what you have to look at is that it was still released in February 2026. So, it's still very recent, this open, not open cloud, but this Hermès agent. And they are already close to 100,000 GitHub stars, I believe, while open cloud, much longer, has around 350,000 GitHub stars currently. So, perhaps these figures are a little bit, are more necessarily up to date, but it still gives you an idea of ​​the traction that there may be in there. And so, it's still quite interesting because, casually, looking at how it was designed and how it works, can even, if you don't want to install agents, help you optimize the way in which you work with your cloud traditionally at home. Typically, one of the huge things that Hermès does is this productized harness engineering. So at some point you have a feedback and feedback loop. They say, if you do the same prompt five times or if you have a sequence of prompts, and so on, write a skill. And one of the great strengths of Hermes is that at the heart of the system, it will observe what you do. And after a certain time, if it sees repetitive actions, if it sees interactions with tools, things like that, it will itself write a skill to help you next time with that. And so that’s something that’s super interesting and that can also help you evolve in your practice. That's why I was saying, when I read and became interested in the subject, I realized that the way in which I could structure the cloud that is on my machine, ultimately, it could really progress thanks to that. So Hermès arrives where, therefore, with the others, with OpenClo and so on, it was necessary to put together almost everything. You had it all. He already arrives with a super well structured harness, which already allows you to do very strong things. And so if we look a little bit, if we break it down a little bit into five layers, the instruction, the constraints, that is to say how you are going to sandbox things, the feedback, how you are going to be able to evolve your system, the memory, since at the heart of all that, there is also the ability to have persistence and orchestration. How do you do a multi-agent, multi-model thing and so on? Hermès comes with super clear answers on this. So on the instructions, typically, when you are working manually, you will write your CloudMD or your AgentMD, well, whatever name you give it, it has its skill system which will auto-create, auto-update. Afterwards, there is this notion of constraint, to what extent you make hooks, you do things which will prevent you from doing certain actions. We saw that one of the things that had been highlighted about OpenClo was that there were people where OpenClo had gone, got carried away, it had started subscribing to additional services, because it had been asked to do things regardless of the way in which it could get there. Finally, there, functioning, there really is a system at the heart of it all. At Hermès, they decided to have a sandboxed permission and execution system. So it’s still really good from that point of view. So, the feedback loop, I talked about it just before, that is to say that where you have to do a manual review, take your prompts a little, put them back in a skills, something, so on, there, automatically, there is a learning loop which is quite automated. It just creates markdown files that you can then edit and modify. But yes, so I didn't look at to what extent, indeed, there was this very manual side of proposing a new thing, and then to what extent he could, behind it, he could have a workflow to be able to activate them or not. I didn't look at that. As for memory, until now you had to serialize the memory a little bit yourself and rewrite in contexts, in specific markdowns, rewrite things like that. There, you have a system, there is even a switch which allows all interactions to persist and which allows you to start interactions on a media. Typically, earlier, I was talking about this thing that was very fashionable, to have a conversation with a... with a... how to say... I'm going to do it with a telegram or a WhatsApp and finish it in the terminal of your machine. Well, for once, this is hyper synchronized since there is a tilt which allows you to have this continuous recording of what is happening. And then, in terms of orchestration, that's also where they made a great effort. This is because you have the possibility of synchronizing tasks, of making multi-model calls, of calling on different models since they are not bound to a specific type of model. We recall that OpenClaw had started and it had a falling out with Anthropik for being very heavily used with Anthropik models. And then Steinberger was recruited by OpenAI. And so, I didn't follow up afterwards, but it's very likely that the following releases of OpenClaw will be very typed things for use with the OpenAI ecosystem. And so, it’s still something. There is really a super interesting feature set from this point of view and which also gives very good ideas if you want to develop an agentic loop or something like that behind it. in particular, there is the memory system which allows us to have things and therefore to also be able to have a retrival latency. If you want to do a RAG, for example, the memory system will be able to really help you since you will be at 10 milliseconds even if you have a volume of documents that you have put in your system which is enormous. So, it's really something that will allow us to build scenarios and resolve lots of different things. And as I said, it's still a big open source alternative to market offers because when we look a little bit, we're listing things and when we look at what's there, for example, in AWS Bedrock Agent Core which is the AWS agentic runtime, we're pretty much on the same feature set. So, it’s really a super interesting system from that point of view. When I dug into this, I really found it super interesting. And so, well, it doesn't happen completely naked and on which you have to build everything since there are overall around a hundred skills which cover lots of different use cases. For example, it's really very, very quick to create something that reads you a list of blogs and gives you a summary every morning to read so you can keep up to date on what's happening in terms of information. And that's something, you'll be able to have it running by itself in a corner on a machine at home and which doesn't require any resources because we'll come to that later. It is very, very well constructed to be able to auto-ostrate models and not be dependent on an external provider. Are there things, Nico, that you have had the opportunity to look at a little, these autonomous agents? No, no more than that. The open-closed part, that is to say they quickly cooled me down in mode where they displayed more security alerts than features at a given moment. So I let it go a bit. It’s true that I also have a little trouble delegating things. I like to control everything. I'm having trouble with this side, just...Hump my back and it'll be fine. I admit that I have a little trouble with this. So I'm learning to relax a little, but not completely. And so, behind closed doors, I quickly let go of the matter. Then since everyone is on it, it's true that I have an inverse relationship. It's more popular. Yes, yes, you're like me. You are like me. You have a certain hatred for things that are too hyped at the moment. That's it. I'll wait until later. And it's true that at this moment, we see a lot of things. It's true that Hermès, we see him quite a bit. I'm going to visit Dolama's blog this morning. He was talking about Open Jarvis who also seems to be a bit in the same spirit. We feel that things are moving. Afterwards, it's true that the use cases, I have a little difficulty saying hey, take all my data and do what you want with it. This is the extreme case. And so, there is still a real subject of everything that is permission. Yeah, you have this sandbox topic which is a real, real, real topic. I remain extremely reluctant. Typically, for example, for things that would interact with my emails or with my Google environment, I use a command line tool called GoCli. It was Nicolas Martignol who showed this, which he had done precisely for his automation around DevOps. He had made quite a few videos on it. And he had actually shown that instead of using cloud stuff, integrated into the cloud, notably in Cowork, to read these emails which are painful, which do not progress, which go at a speed that is completely crazy. There is actually a small binary that can be used on the command line which allows you to output emails and list emails. It outputs that in a JSON format or a markdown, whatever. And then even writing emails and that allows us not to open API access directly to a tool on which we do not control what it will do with it. And there, for once, we are able to say well, wait, this thing, I give it these permissions and so on. And at least I have something I can turn off if something ever goes wrong. And so, it’s still a bit like that subject. So you say yes, you want to control everything. You know what they say. Delegation does not exclude control. Yes, no, but it's the first step and I have a little trouble trusting. Trust does not exclude control. Thank you for taking me back. No worries. No, but it's sure. No, but it’s true that it’s happening little by little. Then we see it with the evolution of the models and then the side... Yeah. But it's true that, for the moment, I can't necessarily find it in the use case. I would have to force myself to do it. I know that indeed, in our professions where we have a need to dissect a quantity of information, to process, even post-process a lot of things. For example, I am firmly convinced that in the production of the podcast, having it go around in a corner and process things, which tries to connect recent news to past episodes and so on, that would make it possible to produce a lot of interesting things. That, I am deeply convinced of. Furthermore, I am also deeply convinced. That was my idea for this episode, before coming across that, it was to come back, to talk again about precisely these tools, these semantic tools, these semantic models and so on, which are very closely linked to an agentic loop which would allow us to have a dialogue with the data, with the data environment. It is clear that these autonomous agents have an interest in being able to process this. We also say, well there you go, we say that agents are things capable of working on boring tasks, those that we, humans, no longer want to do at all and so on. Ok, well the thing in data on which we are still a little bit uncomfortable is the subject of documentation, of validating, of doing a bit of linter work on the data environments, of validating that we have put the comments on all the columns, of validating all these things, finally in quotes that we try to do in CI-CDs classically in code, I know that on data it is more tedious and I remain convinced that this, for once, agents who are circling in a corner, we would be able to do this data stewardship work quite well, in fact. So. Oh yes, I find that there are plenty of uses, after that it's more about a personal age, for the moment, I can't do it. The news at the moment when everyone, we feel that there is tension over the use of tokens, that is to say that that... Yeah, yeah. It's a good time to test which was a few months ago where you, as it was in open bar mode on your cloud subscriptions, OpenAI and others, it was easy. There, when we saw that you are going to pay with the token, it risks being a little less... So, that requires having a machine with a little power, that is, but if you are not doing interactive, if it is agents to whom you give the time to think and to be able to provide a response, typically, we will come to it just after, the variety of models, even with quantified models that you can run on your machine, it is still quite important, that is, since there is indeed the possibility of using local runtimes and possibly delegating some one-off tasks which would require a lot, a lot of diversity of performance and particular features, to send that to a remote model. But I remain convinced, the interest of these agents, where I never bought into the OpenClo story, was, yeah, ok, if it's to run something on a machine at home and in the end, eat the token from an external provider, what's the point, what? What's the point? So, at a given moment, I see the real interest in the subject of how I can run things entirely, locally and possibly say that by extension, you can have things where you also guarantee your sovereignty. For me, it still remains a subject on which I am super cautious, super vigilant and therefore, in quotes, I prefer to have a deep-sic model or a quen or a model like that which runs locally than to send everything outside. Yes, no, but the same, you just have to have the machine. You just have to have the machine. You just have to run a machine. And indeed, in the heat we are currently having, our offices are already overheated and therefore, it is indeed perhaps not the best thing to do. we completely agree. So, we often draw a parallel between Hermès and OpenClaw. We've been doing this for a while now. These are projects that are very simultaneous since, as I said, OpenClaw which was called CloudBot and then MoldBot made by Peter Steinberger. Is it the end of last year? Yes, that's it. It launched in November-December 2025. It went viral in January-February 2026. And today, Hermès Agent arrived in roughly the same time frame a little bit later. And so, where there is one that is simpler, OpenClaw, ultimately, it's not much. It's just a Control Plane. You have your agentic loop and then that's all there is and you integrate lots of things around it. Where Hermes is much more integrated and provides everything and in particular this mechanism that they call the learning loop and persistent memory. and so, it's really a gateway on one side and a full agent runtime on the other. However, it's always quite funny when there is competition between different tools in a market. There is a command called Hermès Claw Migrate which allows you to retrieve what you have already done in OpenClaw to switch. So, this kind of thing is always quite funny. What... We talked a little bit about this security. Hermès, by default, the architecture is designed to be safe. These are not a posteriori patches that allow you to correct CVEs that could be there. So, at a given moment, there is really a noticeable difference and that is perhaps what will make us more inclined to call it, to use it than to use its predecessor. So there really is this important point. And there is really one point that is super important which is the fact that Hermès is model agnostic. and that was still the first thing I looked at. It's the extent to which you are free to do what you want with it and use it how you want. So typically, you can obviously use cloud providers. So, it's going to be Noussportal to use the Hermès models that they use, that they offer. You can use your Anthropik account. So, we will come back knowing that this will change in the coming days since I believe that on June 15, Anthropik will roll out this new billing method since until now, precisely, OpenClaw is one of those things which has meant that the profitability of Anthropik in particular has been seriously undermined since remember at the start, you could have your subscription at 20 dollars and make API calls on the basis of this practically unlimited subscription at 20 dollars. Let's say it, I say it practically, but you managed to type quotas from time to time, but it was very very distant and in particular all the non-interactive modes of cloud were possible by the API call, by the cloud-p command which allows you to pass as a parameter in a line cloud command to pass a prompt and retrieve the output. So, all this stuff was OpenBar and so, Anthropique said no, wait, we have customers who pay 20 dollars and who cost us 1000 per month, so that's enough, we're going to stop there and so, from June 16, normally, there will be an API credit envelope which will be provided for each of the plans that you take and therefore, you will have this envelope and once you have exceeded this envelope, on the other hand, we will be paid to the token of API calls and therefore, it will still be another matter. So there, we're going to realize that people are going to say to themselves yeah, if I can use less efficient models and run them at home, it's still going to be pretty good and so, that's where, indeed, Hermès, they were smart, they said yeah, wait, from the outset, we, you can use behind providers with API key or even self-hosted, that is to say you put an Olamas or an OMLX, the framework from Apple and you run models locally and that's what you're going to use as an engine and eventually, you can do routing to an external engine, that is to say that typically, if you have something that requires the power of a cloud opus 4.8, well, let's go, here we go, you make calls when it's necessary and so, it's still super important from that point of view, this ability to look for broader things and be able to have something extremely open. So, for me, that's really what interested me in this and so, if we look a little bit, I have a little bit, as I said, I dug into the thing a little bit and so, I looked, if you go to Hermès Agent, you will be able to see, they have in particular guides and tutorials where you have precisely explained how to do, for example, a daily briefing bot and this daily briefing bot, he, well, you go into the thing, say, there you go, it there is a cron scheduler which will start the bot at one o'clock in the morning and therefore, it will do web research, it will do a summarization of the content and clearly, if you have, it's really a few minutes to do that, what, that's where we say to ourselves, ultimately, we all perhaps have repetitive tasks that at a given moment, we will have to go, we will have to, we will be able to operate like that and have something that runs quite autonomously therefore, it's still a matter of time afterwards, it's a matter of time to try things but I'm quite... And earlier, we were talking about IWS, you see, you can even go and integrate your Hermès Agent with Bedrock since you have the possibility of using your IWS Bedrock account to be able to run this agent so, clearly, I think that these are things that we should still, that we will have to look at quite closely to be able, quite closely, sorry, to be able to create things. And so, we were just talking about the Telegram Assistant pattern, it's obviously something that they have documented where you create your bot, you create your new bot inside Telegram and you get an ID and then you configure that in the gateway and therefore, you will be able to communicate with this Telegram. So. And closer for us and for the developers, obviously, earlier, we were looking for what could be the good use cases and indeed, the peer review on GitHub is indeed something which again can be something interesting to do, to say, well there you go, you put a GitHub CLI on your, where you run your Hermes Agent, you do it and you ask it to do the reviews and summarize the peer reviews for you to be able to easily do that. So there you have it. Afterwards, as the commits are made by agents now, if we have the peering done by agents, I don't know anymore, I don't really know if we won't have at some point a total lack of control over everything that happens. So it's perhaps... You have a real subject and we feel, annoyingly, that it's really the counterpart of the moment which is to say ok, the agents, finally, the AI tools can do lots of things to us afterwards, that's what the... What level of control do you keep on the subject and also in terms of validation and in terms of knowledge of the model, finally, the model of the world which is represented by your application and it's true that it's... It's tempting to do, well come on, it's okay, I'll let it happen, it'll go well, I'll quickly look at it and then gradually you feel that you're losing control a little and it's true that you have to find good compromises. Yeah, we definitely agree, we definitely agree that at a given moment, just now, as you said earlier, if you actually let him recreate skills, at a given moment, you have a real danger in the fact that you know... At the beginning, you're going to check and then afterwards, the thing runs in a loop on its own. So, it's clear that at some point, you're going to end up with something that can end up being dangerous. To summarize. To summarize. But you, I imagine, you, in what you're building in terms of monitoring agricultural infrastructure and so on, you probably have tons of monitoring tasks for things like that where you have to go and look frequently and so on, on which these are things that are super important. I see with a client, they have a terminology which is keep the light on and therefore, it is a big subject which occupies a lot of people every day to go around lots of assets to check that it is good. It is a complex monitoring which cannot be put in a simple dashboard which requires a lot of actions to check in order to do something else and that, I think that these are things where in the end, we will end up being able to get there and for agencies and tasks that are still tedious and with little added value which can really make a difference in my opinion. The limit of that for the moment is that all that is greenhouse equipment is often automatons which are offline and today, finally, it is rather a bit of old technology. We are not so much on the operational part, we are no longer on the analysis part so we extract the data, we recover them, we grind them and then we possibly send back instructions and then we recover after the... But we are not really at the level of the agent who... We can manage them but we are not really on the purely operational monitoring part. So it's true that we're not completely at that level and for the moment, it's more... Well, when we see the interfaces, it's more old stuff. It's like in industry. In the industrial sector, it's all to nothing. That is to say that extremely recent machines will be hyper connected, will be capable of being integrated with management systems in... But... Old machines, you have in the flora of... Industrial, things that will never be connected. So for which it is indeed very complicated. And then afterward, you also have your serfs who are in certain places. I'm not even sure if they have internet connectivity. That's it, yes. No, but what I mean by that... But yes, no, but... With a flow rate that would allow for an interaction we continue with it. Yes, after that, it’s not necessarily a permanent connection. It's true that there, we have certain projects where it's... We have holes in the... Well, which fill themselves up afterwards, but it's not necessarily real time. So there you have it. Yeah, yeah. But yes, in good potential. Yeah, yeah, absolutely. We can... I think we took a little look at this Hermès Agent. I'm sure we'll talk about it again. We have brought up some little news there, recently in the... In the Slack and in particular... In particular the story of... We can emphasize this because in addition, it is a Frenchman who wrote a LinkedIn post on this. It's Steve Morin who you can also listen to in this excellent French-style podcast that I love listening to. And it actually talks about CUDA and where... I don't know if it was you who didn't bring this up, Nico, in the Slack. In fact, to provide some context, Steve Morin is the originator of the ZML box which is ultimately the idea is to have a layer of what I understand to be a somewhat generic attraction layer to allow the models to be run regardless of the underlying hardware. And it's true that today, well, indeed, the elephant in the room is Nvidia's Cuda which allows... which is a bit of the de facto standard in this world. And it's true that the more things go, the more we actually see that the projects which go so far as to finally free themselves from Cuda which... But even from... I don't know, I listened to a podcast for a long time where Steve Morin said in Cuda there are lots of good things and then there are things that are less pretty. It's an old Cuda, we must not lose sight of the fact that it dates from 20 years ago. Yes. Not far. Version 1.0 of Python has just been released but it was announced yesterday, I think I'm lying. Yeah, yeah. But suddenly, there were things that were specific to Nvidia but it's true that Nvidia had the strength to bring both the hardware and the software and as a result it's a bit the de facto standard and it's true that there with... notably in the work of ZML and I don't know, I think he started... Michel Kardec who also... or no, I don't remember who... in short, who had highlighted other projects but there are plenty of projects which ultimately go towards a slightly more generic side which could allow us to free ourselves from Nvidia and I think that there is still a big threat for Nvidia's monopoly which is that one day we can manage to do without art and no longer do all this part which is a bit low level... that's a big big threat for them. Nvidia is threatened on two levels, that is to say that a first level which actually this software stack on which they were hegemonic they will perhaps be less and less first point then second point more and more still even if it is not yet in quotes in production everywhere alternatives to Nvidia technology for not necessarily for the training of models where it will probably remain for a very long time this brute force that we can have will still be very valid but there are still quite a few of alternatives which are in the process of rising on the inference and moreover even Nvidia in quotes is working on this alternative they bought this company which is called Grock nothing to do with the Grock it's an exclusive partnership ah sorry I was wrong sorry I was wrong I apologize it's an exclusive partnership which is to show good will very linked to the exchange of shares that's it ok officially it's not purchased it's a partnership I I don't know what exclusive I don't know if as we said I think in a previous podcast I think that the only people targeted by the maneuver were the anti-trust commission so basically we can see that there is something that is shifting and ultimately it is perhaps a very good thing because when we see deploying a use case based on LLM and others the infrastructure cost that you have to take to be able to have access in particular to inference based on NVIDIA H200 or H100 even H100 it is insane so the fact of having both a software stack which allows it to be done more easily and having hardware mechanisms which are simpler than this big artillery is clearly something which does not go in the direction of the development of NVIDIA it is clear another article which amused me I read but I find it interesting who is on I believe it is people from Uber who made it f have a bit of a sort of feedback on the identity crisis ultimately of the agents and it is true that there is a real question on this since in the measures where we delegate the response that they make I believed a little I did not go into the detail of their response it seems to me a bit of a gas factory but in any case the problem is interesting which is to say finally when we entrust a task to an agent we lose a little for whom we are doing it we lose this whole notion of traceability authorization of all these subjects and it is true that suddenly in their can it actually lose context whether for debug support or anything or also to have hey if I have Vincent who asks something to an agent we must know that it is Vincent who has the origin of the initial request and as the agents in addition can chain each other we must ultimately know all this all this traceability all this chain of operation to possibly deduce I don't know anything about permissions authorizations whatever we want and it's true that there is a real subject on this that we don't necessarily think about we say to ourselves it's good there is an agent we ask him something and then it will go well but it's true that the subject of this whole notion of identity takes up a real subject with agents yeah yeah yeah and well that's always this story of impersonation of finally we've had this subject since the dawn of time finally to know how to trace when we make multiple epia calls and finally know who is at the origin and there we end up with the same thing you are right when you speak to an agent who calls on sub-agents of the sub-agents of the sub-agents who each have their level of permission and so on what do we do do we carry around the token which could have been put in the first finally the token I mean the security token the token which could have been put in the first request and do we take it from start to finish or do we reset that it is indeed and moreover we can clearly see that there is a small movement that is taking place on the on the place finally how we make the link with the human who revolves around these processes and I had found another article done by Adi Osmani who said that ultimately the bottleneck in all agentic processes was going to be the human since at a given moment there is no that a single person who is capable of judging merger to arbitrate what the agents produce is a human brain and therefore you could deploy 250 Mac Minis in your office to do all the work and so on on both sides at a given moment everything will have to go through a single little pipe it's that of your brain so so it's quite amusing this thing what it's not the exchange the tax the orchestration tax and it's just it's true that it puts a lot of good advice we're going to say to try to protect ourselves from it but here the article is nice yeah the article is nice and well Nico I think that we were both concise and and we were able to talk about lots of things so it's it's nice it's not bad it's for me a little plot and then some news around it's not bad I like it yeah I like a subject on which we spend a little time and then we try to review two or three more things it's still not bad and well listen it was cool to record this episode dear listeners dear listeners we hope you had a good time listening to us as good a time as we had recording it and we'll see you again very soon very soon dear listeners Subtitling Société Radio-Canada