Hermes Agents Now Manage Their Own Work

0:00Hermes Agent just shipped another update that is putting a lot more pressure on the Open Claw community and this isn't because Hermes just magically is winning the race now. That's not the point here. They're two different builds building for different audiences and the point of this is basically Hermes is moving from viral agent project. I mean it's gone crazy over these last few weeks to an actual AI operating system and you're hearing more and more AI operating system more than ever than you've ever heard before. That's what we're moving towards and that's what Hermes is building towards. So basically faster startup, better memory, of course that's in there. But real multi-agent Kanban boards and prompt injection defenses already built up out of the gate. They

0:46have secret management skills and they have over a hundred preloaded skills inside of Hermes Agent, which is pretty crazy. So they just made this release a couple days ago. I'm going to be breaking down why it even matters for people that are using Hermes. So basically they got a lot more sophisticated in this release compared to what Open Claw has been shipping lately. You know, I'm going to be talking about the new point 15. It's crazy we're not even on 1.0 yet, but this release does matter specifically for Open Claw builders too because I see a lot of people switching over or also adding Hermes Agent in addition to running Open Claw. I think people are still getting caught up in the hype really and instead of actually building workflows, like

1:31figuring out what to choose, like Open Claw or Hermes. You need to make the decision and you need to do it now because you're losing time. So this official release page, you know, is called the velocity release. It was published like I said couple days ago. The release notes say, you know, since the latest version Hermes had 1300 commits and 747 pull requests. So there was about 1700 files changed in this last update. So, these numbers matter because this is not one tiny feature drop. This is a pretty giant update and a step in the right direction and it's a huge velocity signal. So, when a rival agent project to Open Claw ships that much code, you know, we shouldn't dismiss it. We need to pay attention.

2:16So, we need to actually study this. You know, I'm saying that because I'm someone who uses Open Claw. I use it every single day and I love Open Claw. You know, I've spent a lot of time building my whole workflow out inside of it. I'm not going to push it aside, but I have been using Hermes a lot more specifically for marketing. And, you know, Open Claw, like I said, runs a lot of the content back end. It does a lot of the podcast outreach, you know, my database workflow and all that stuff. So, All right, real quick, I want to talk about our community and what we are doing right now specifically. We do four live calls a week and we are in the middle of a 24-week sprint where we cover what are cron jobs, what are scheduled agents, building your first Open Claw automation, building a

3:01coaching follow-up workflow, building a customer support work agent. We talk about local LLMs, model size and quantization, embeddings, AI security basics, evals, agent architecture. This is what we're in the middle of right now. So, if you want to learn about prompting and context engineering, software basics, APIs, what is Supabase, what is GitHub, what is Vercel, how to use Codex, Claude code, Open Claw, Hermes, how to build a content agent, how to automate content, how to get your first customers for the app that you built. We go over security, evals, embeddings, inference. We go over what is inference engineering, what is AI engineering. So, we have four other coaches in the group. We do four live calls a week and we are building out this entire course content to be able to

3:47help people where they're at. They can learn on their own time. So, if you want to learn about AI and get better and level up your AI skills, whether it's building or marketing or learning and hosting local models, check the link down below, come join and come build with us. We have over 150 builders in there and we'll see in the next one and let's get back into the episode. Doesn't mean I ignore Hermes, it just means that I'm really so tied up with open claw, I'm not going to let it go. So, you can see here the big refactor, you know, is no longer the run agent Python script is no longer 16,000 lines of code. That's pretty crazy. The can ban grew into a real multi-agent platform, 104 pull requests end-to-end. They have a cold start that keeps the wave going so much

4:33faster on the startup. The prompt wear defense got a lot better. So, brain worm class attacks blocked at three checkpoints, Bitwarden secret manager, skills bundles, like I said, they have over 100 included skills just on an install. So, there's a lot to go into here and, you know, I'm going to break down what is exactly important. Like like the first thing that jumped out to me specifically in this refactor is the run agent Python file. Like 16,000 lines down to 3,800 lines. Like, what exactly happened there? Um that is a 76% reduction. The code that was extracted into basically 14 cohesive modules under the agent folder while keeping compatibility through thin forwarders on the AI agent. That's

5:19crazy. Like, this is the kind of change that does not make a flashy thumbnail or anything, you know, I'm not going to put that in the title, but this does matter. And the agent loop is a giant 16,000 line file. Every contributor has to crawl through and mud through it and change everything. And, you know, when that gets split into modules, the project gets easier just to understand. Like, easier to patch, easier to review, and easier to extend. You know, that's the first builder lesson here. Another one that sticks out is the OpenAI style execution guide applied to Grok models. I use a ton of the Grok and X API calls inside of my Open Claw just to help do research. It is a huge tool to help me find trends, to also make a lot of the

6:05content, to see what even is worth making. And that should also be implemented inside of your Hermes agent 100%. If you're not doing that, it's pay per use now. You don't have to pay to $200 a month for the X API. Like if your agent system is working, but you can't understand it, you don't have a system. I'm sorry to say. You need to understand what's going on in the back end, what workflows and what tools your agents use. You know, the the second big thing also, too, I'm a big fan of Kanban boards. I don't know about you, but the Kanban board grew into a real multi-agent platform across 104 pull requests. Like what did they exactly do there? The release notes basically say it now supports task auto decomposition, swarm typology, and also scheduled

6:50tasks. So a work tree per task, per task model auto overrides, and then board level default work directories. So that's a lot, I know, but here's the plain English version. Hermes is just trying to make the agent work look less like one chat window and more like a real job board. This is a great way for agents to stay on task, especially if you have an or- chestrator agent and an a worker agent. So the worker agent takes the work and the orchestrator agent checks the work. And that feedback loop is going to help you actually have workflows that work in one shot rather than having a workflow break mid-run, right? So basically what Hermes is doing with the Kanban board is like a task will come in, it gets broken up into

7:36smaller tasks, and then agent workers can claim them. So some workers can use cheaper models like a Llama or Qwen or if you're a local person or maybe just a lower end GPT or a Haiku model. You know, and harder tasks can use the better models like GPT 5.5 or Opus 4.8. Each task can have its own work tree even. The system can also detect stale workers. So, after that, a verified worker and a synthesizer can pull it all back together and then report back to that orchestrator agent. This is the right direction and in the right workflow for AI agents in general. The future of agents is not one perfect chatbot. We are so past that. We don't want to one chatbot that does everything. It needs to be broken up.

8:21You need to have it a managed workflow. Because at some point, if one chatbot does everything, what if one thing breaks? Then the whole system fails. The future basically is supervised work systems. So, that Kanban board is an amazing setup. You give the system a goal and then it breaks that goal down and it assigns the work. It tests the result, it then it retries when something breaks, and then it keeps the receipts and that feedback loop keeps going. That's what normal businesses need. So, you can see here the agent loop conversation, it breaks it all down. I have this link down below. You really need to just check it out or give it to your agent. This was a huge one here and that's why I wanted to make a video about it just to keep the Hermes people up to date. It's pretty interesting on where the direction is going. Most people in most businesses don't need a chatbot at all anyways.

9:08They don't need a chatbot that writes code. They just need a worker that can prove what changed, why it changed, and where it failed. And then what still needs review. And remember, think about boring tasks. That's the play here cuz the money is in the boring stuff. I'm sorry to say. Hermes also says that this release cut the per conversation function calls by 47%. So, well, if we're talking numbers here, that's from 400,000 to 213,000 in a 31 turn chat. So, it says Termux cold start dropped also about 3 seconds. That's what it used to be, 2.8 seconds. It also says Hermes version cold dropped to 701 milliseconds down to 258. You probably won't notice it too much, but that is a big difference if you're

9:53running multiple Hermes agents. It does really sound small until you actually use the tools every day, you would probably notice. And setup time does matter. I mean, Open Cloak can take a couple minutes just to restart the gateway. It's pretty crazy. The difference between an agent that, you know, wakes up in a second and an agent feels like it's dragging is really difficult. And it's it's annoying, honestly. It changes basically whether you keep using it or not. For a lot of people that are using these agents every day, that does make the difference. Normal people really don't care about that. They don't care about your specific architecture diagram. They care The only thing they care about is whether they can send the bot a message and get moving. This is where Hermes has been smart the whole time. The readme actually leans hard into normal builder

10:38distribution. It says Hermes can run on a $5 VPS. While I don't suggest it, you can. You know, it can even run a GPT or a GPU cluster or a serverless infrastructure that costs basically nothing when idle. And it says also you can talk to it from Telegram while it works on a cloud virtual machine. It supports providers like News Portal, Open Router, Nvidia NIM, and OpenAI Hugging Face. There's a lot of custom endpoints, too. That is exactly the kind of positioning that gets non-technical builders more interested. And I feel like that's the play here. Hermes is so easy to set up, like my dad can do it. Not because they want actual architecture, but, you know, automating a workflow seems interesting to them. They want a worker that is not trapped

11:23on one laptop, too. You know, if you want to be able to access your Hermes in the cloud, you can do that easily. Fourth [clears throat] thing in this release, I'll have to say, is the session search. Hermes says the session search was rebuilt completely, so it no longer uses the auxiliary large language model, and it no longer costs 30 cents per call, and it no longer takes 30 seconds to summarize three sessions. So, that's what the release notes say, and that's what the discovery is now about. It's about 20 milliseconds instead of 90 seconds, which is crazy fast. So, they call it 4,500 times faster and free, which I'll take any day, and that's a huge one. Memory is probably one of the most overrated and underrated parts of agents. I mean,

12:08there's people that spend weeks, like, stressing so much about a memory, and then there's people that don't even think about it. So, I think there's a fine line that's like, you're kind of over-engineering the memory part, but also pay attention to it, but just don't spend too much time on it. It's not like, you get it dialed in. Like, it's important, but don't spend your whole life trying to figure it out. You know, it's overrated in the sense when people treat it like magic. Like, it's not perfect, and then it's underrated when builders realize the whole product that you're building breaks, basically, if your agent can't find the thing it already learned about yesterday, which is probably the most annoying part, specifically with Openclaw. And I know if you use Openclaw, you know exactly what I'm saying. Like, what a good memory does and what good memory search does, and this doesn't sound fancy, but

12:54it just needs to be fast, cheap, and boring enough to call all the time. Like, I use Obsidian Vault, and I use Obsidian Vault for my Openclaw, my Hermes, my Claude code, and my Codex. They all can read what I'm working on. It's boring, simple, but it's effective, and that is why it matters, you know? Like, if your recall is slow, the agent will skip it. If your recall is expensive, you're going to ration it because you're not going to justify the costs. If recall hallucinates because it's basically summarizing the wrong sessions, you're going to stop trusting it, and then it's going to feed bad context into your prompts. Hermes is attacking that problem specifically here and I love that. The fifth thing too, people are going to overlook this, is security. I know we've kind of banged that drum for a long time now, but

13:40security in this release notes say where Hermes added prompt where defense against brainworm and the class attacks. What does that it mean? Basically, it scans recalled memory at load time. So, it marks tools results with a delimiter markers. So, basically malicious files or remote services cannot impersonate Hermes system content and it also adds Hermes threat patterns for command and control style injection, which is super important. We don't want to get prompt injected. This is one of those things that sounds really stupid and boring until your agent reads a web page or ingest a poison instruction and then starts acting like a crazy person, right? That's not good. If you are building agents that read email or even

14:25just browse the web at all, like Fire Qual or Brave API, that's what you're using, or as if it's processing documentation or searching memory or installing skills, prompt injection can happen. But, the job here, that what they build in, that's what it's meant to block. And the more useful your agent becomes, the more untrusted text it touches. So, the more you're relying on it. This is built into the latest update, which is awesome and I like seeing Hermes ship defense here. Not because I think, you know, any one of these releases solves the problem forever, but because it tells me they understand where the real risk is. So, if you see here in the release notes, it's like number, yeah, it's like number five, prompt defense. And you can read more about it here. But, it says Hermes now defends the context window against

15:11prompt injection. Attacks that try to hack the agent via tool output, recalled memory or sorted skills. It refers back to the single source of truth with 15 new brain worm C2 patterns and then recalled memory is scanned at load time. Tool results get delimiter markers, so a malicious file or remote service can't impersonate Hermes own system content. So, that's pretty cool. I love that. They go into other security updates here, but I think that's one of the more than important ones. Also too, you know, the sixth thing is the secrets. Hermes added the Bitwarden Secrets Manager support. You can see here. Basically [clears throat] says one bootstrap token replaces every per provider API key. Stop keeping plain

15:56text API keys in Hermes environment variables. Install the Bitwarden Secrets Manager. So, basically the release says one bootstrap token can replace a bunch of API keys credentials. So, that could come from Bitwarden at startup and key rotation in Bitwarden can take effect because Bitwarden defaults to the actual source of truth. Again, this is not a crazy like going to change your life update, but it is important for, you know, if you're using Hermes, you want to know these updates, right? You want to know how is this going to impact your workflow? This is what real operators need. Most agent demos are, you know, held together with local environment variable files and pasted API keys and hope. That's dangerous and that is fine for like a weekend build, but when

16:42you're pushing to production, you need to be much safer. And it's not fine when an agent is connected to your business and your customers and your cloud accounts and your database and your calendar and your posting tools and and all that payment stack. It's scary, especially if you're dealing with customer financial credentials. So, secret management is not a nice to have. Once agents can do real work, this is necessary. So, I'm going to talk about Open Claw and and I feel like Open Claw, you know, it has incredible infrastructure around sessions and tools and channel delivery, memory and skills and all the native Open Claw workflows, but Hermes is pushing hard on that self-improving skills in cloud first distribution, cross-platform messaging, a ton of model flexibility, the Kanban board, and also memory search. And now it's velocity. So, if you are a builder

17:27right now, you should not be emotionally attached to Open Claw or either one. Just use them. Just start using the workflow, start using the agents, start automating something. I mean, I made a video yesterday automating our entire pipeline and content pipeline using Claude code and Hermes agent. So, just steal the patterns that are going on here. You know, here's what I would copy from Hermes right now if I was building is copy the job board idea. I love that idea. Don't just tell an agent to build something and trust the final paragraph. Force the work into task. You know, create an HDMI or HTML Kanban board where your agents can manage the work. Enforce each task to have a status card. Force each worker to actually leave evidence that it completed the task and

18:12get it done. The second thing is copying the per task model routing. Like you need to be using cheaper models for boiler plate plate and expensive models for the hard judgment. Like keep the low-end models that do the web scraping. You don't need to be using Opus 48 for that. That's what I would be doing. Those are the two first things that I would be implementing right now in my workflows. Like after Hermes 0.15 shipped, you know, just the last few days ago, there is a lot of fuss about what am I going to be using, Open Claw or Hermes. Hermes has been blowing up. If you see on Open Router, the most used agent is Hermes by 5x over Open Claw. So, people are, you know, using it as real software and that's how it should be. You know, you ship and things break naturally and you

18:59patch them and you keep going and that's what the Hermes and Open Claw community is doing. But here's the biggest mistake that, you know, agents basically keep making is you think and agents think that your stack is perfect because the demo worked once. That's just preconceived notion. Like it's not going to work like the demo worked. We're all building different workflows. You know, and what I'm even thinking, you know, for you is like how builders should even use this specifically. Like what do you do with this? Like if you're using Open Claw, keep using it. Open Claw is the operator story and it's still strong, especially if you spent time invested like I have. But you really need to test Hermes. Like you need to get it going. Install it in a sandbox if you're scared. Run the same workflow you basically run in Open Claw

19:44and give it a real task. Like don't give it a toy prompt. Just like something messy. Something with files, memory, tools, and follow-up. Something like that. And then compare the three outputs. Like how fast did it start? How well does it remember? And does it actually prove the work? That's what I would do. Um you know, not Twitter hype. I love demos and everything. I I I'm sure you do, too. They're very cool. But when you actually use things in production, you're like, "This is way different than what I saw in the demo." So, like I said, you know, Hermes has 175,000 stars. Uh about half of what Open Claw has on GitHub. And I think stars are more like of a vanity metric right now. They used to be super important years ago, but now it's like anyone can star a project. So, I don't know. But still, stars are a signal. Um

20:29receipts are the product. Things are being built on Hermes. If you go to their website, you can see actual use cases. I think this update is one of the most important Hermes releases I've seen because it is less about the shiny tricks, but more about the boring pieces that make it more reliable and make your AI agents work. Speed, memory, task orchestration, security, secrets, and messaging. Like there's a lot here. There's a lot. It's overwhelming. That's why I didn't cover it all. But, you know, I would say like give this to your agent and tell it to summarize it for you and see how it can affect your workflow and what they can implement in your workflow today. You know, that is the stack here. And if you're building with agents, this is the level you need to be studying at. Every release you need to be looking at and paying attention, especially if you have agents

21:14running. Do not just ask which model is the smartest, like stop that. The models are all smart, they're all really good. Just pick one. You need to be asking which system can keep working when the task gets messy. You know, the first worker will fail, and the memory is buried, the API key rotates, the web page is hostile, and the result needs proof. That is where agent systems are going, and Hermes just made this race super interesting. I see a ton of people leaving open claw and switching to Hermes. If you want help actually building real agent workflows like this, not just playing with prompts, come join the Shipping School community, that's exactly what we do. We do four live calls a week, and now we are focusing on more tactical tutorial-based courses. So, specifically, installing Codex, how to get your first Codex

21:59automation running, installing Claude code, how to start building an AI agent workforce, how to actually use AI agents. What is AI? What is machine learning? We also teach about how to install local LLMs on your hardware to actually be able to cut down on costs. So, we have a lot planned for the actual community. We are in the middle of a 24-week sprint, and inside there we're covering all the basics from AI agents to machine learning and getting local LLMs installed. I'll put the link to the community down below, check it out if you're interested in learning about AI. And what do you think about this new Hermes update? Leave it in the comments below. Have you ditched your open claw update yet, or have you ditched your open claw setup yet? I haven't, but maybe you have. So, I want to hear back.

22:46We'll see you guys in the next one. Have a blessed day. Bye-bye.