Hermes Agent Update v0.13 is VERY POWERFUL! (Tenacity Release)

追加: 2026-05-11

[00:00:00]All right, folks. Hermes Agent just dropped another update V0.13.

[00:00:05]This is the tenacity release. This is a very exciting one because this pertains to one of my favorite new features of Hermes Agent, which is the multi-agent Kanban. Actually, we've been doing several setup guides on how I've been using Kanban to improve the pipeline of my projects. In fact, yesterday we did a updated Kanban guide, which was actually from the V0.13 release notes. So, we sort of front-ran before they updated it. But, yeah, all in all, V0.13 covers a lot more than just multi-agent Kanban. So, we're going to briefly go over what's new with the multi-agent Kanban. If you haven't watched our previous videos or if this is your first time on our channel, we're also going to dive a bit deeper with the new {slash} command {slash} goal, which is sort of like the Ralph loop that ensures your agent stays on track, which is a very helpful feature, as well as the new video analyze tool.

[00:01:10]This is also super super helpful if you need videos for reference to give to your agent to build something. All right. So, we're going to cover all of that in this video.

[00:01:22]There's also other things that you need to be aware of. We actually tagged the security wave of the V0.13 as critical.

[00:01:30]So, you need to be paying attention to this so that nothing gets exposed or nothing breaks in your Hermes Agent setup. Now, if you've been following us for a while, you know that I am a huge fan of the Hermes Kanban. It just makes your life so much easier. Any pipeline you have for your project can be significantly improved because of the actual parallel execution of the multi-agents that you create uh depending on, you know, how how complex the pipelines are for your project.

[00:02:05]So, we did a full uh dive into these setup. So, yesterday we were talking about uh setting up multiple boards as well as uh config- configuring the.env for the profiles that you spin up. Uh and then here we also did one um pairing Kanban with Cron Job. However, we we still notice from this release uh there's no native synergy between Kanban and Cron Job. Uh my Hermes agent had to create uh its own deduplication logic for my Cron Jobs to work uh using Kanban feature. So, I hope that the Hermes agent team considers this uh and then makes a natural synergy that way you don't have to create your own script to make uh Cron Jobs with Kanban uh work together. So, you can watch these videos uh if you want a full step-by-step guide for this. But, the important thing here to note about the uh multi-agent Kanban from this release note, we get to learn a bit more on how durable the Kanban can be. So, basically under the hood, there is a heartbeat monitoring that now detects uh if a worker process dies unexpectedly. So, that's actually one of the key features of Hermes uh Kanban workers is they would keep trying, right? If a if a task failed, they would keep trying and you can see in the run history how many runs in total did they have uh done to achieve your task or to complete your uh their work. One of my agents actually did 81 runs just to finish their research cuz it kept failing with the browser automation until it figured out how to use the browser automation correctly.

[00:03:45]Uh but, there are in some cases where they would try try try try uh but, they It not report to you if it failed or not. So, this heartbeat monitoring would would tell you if they died unexpectedly.

[00:03:57]Then you also have a reclaim logic that would grab those tasks back and then recuse them. So, this is the solution for workers that would die unexpectedly.

[00:04:06]And then we also have a new function being introduced here, which is the per task retry overrides. Uh it basically lets you set different retry policies for different tasks. Uh it's pretty much very similar, I'd say, to a {slash} steer. Okay, now moving on to the {slash} goal command. Uh it's super super helpful because most problems for I think a lot of you guys would be, let's say you start a session and say something like, you know, build a trading bot uh with Discord alerts. And then the agent would start working uh and then 10 turns later, they would have uh a webhook error that they need to debug, right? Then 20 turns later, they're going to tweak the Engine X configs for the webhook endpoint. And then around turn 30 is when they realize, "Wait a minute, you know, what was I building again, right?" So, instead of being a developer to build a trading bot for you, by turn 30, they just became a webhook debugging assistant. So, that's uh context drift.

[00:05:10]So, in long sessions, especially with context compression kicking in, the original objective gets pushed out. So, the agent is still helpful. They still give you good answers, but they are not focusing on helping you what you originally asked for. But by starting off your prompt with {slash} goal build a trading bot with Discord alerts, you get the a persistent cross-turn goal lock. So, you set it once and it survives compression. So, um you know, you should get in the habit of hitting {slash} compress and then compress what's important uh the goal, right? But most for most people maybe the sessions might go on too long and they would auto compress. So, you don't get to have the choice to compress which information you want. So, the {slash} goal just um makes it persistent so that the goal is all always survives uh every compression. And in fact, you it doesn't have to be the beginning of a session.

[00:06:11]You can even do this in the middle of the session. So, when I was trying to fix the cron job uh Kanban this morning because of the shell date date syntax, it asked me the same clarifying question six times. I was like, "What's going on?"

[00:06:27]Right? I even said, "Why are you asking me this six times?" And he's like, "Oh, you're sorry. I'm sorry. You're right.

[00:06:32]Let me check it again."

[00:06:34]And then it asked me again a seventh time. So, I closed the session, type uh {slash} uh updated Hermes first, and then {slash} goal, and then it remembered, "Oh, yeah. Right. I should not get stuck in this loop because the cache session was already deleted, but it was attempting to do something impossible. So, that's why it was stuck in that loop." So, what happened was it fixated on becoming this pro problem solver for this impossible task. Whereas the goal should have been, you know, "Give me my report today." Right? So, the {slash} goal can really help out a lot and you can use this in the beginning, you can use this in the middle if something really becomes frustrating for you. And I'd say it's super helpful if your agents have to use a lot of tool calls. If they have to interact with a lot of API endpoints, this is especially useful. Okay, you don't have to use this for simple tasks, but if there are API endpoints or tool calls, multiple tool calls, you know, go for {slash} goal.

[00:07:36]Okay, next up is the video analyze tool.

[00:07:38]This one super super helpful. You can just give a YouTube link or a URL with a video file to your Hermes agent and get them to analyze the video. You can also upload an MP4 file in your workspace directory, point your agent to that file and then tell them to analyze the video. But there is a caveat here. I think so far it only works with Gemini models. I'm using Kimi 2.6, which of course is compatible with video and image, but I tried it, but it said "Error, model does not support video input." So if you're using Gemini model, this would work. But if you are using some other models, but you don't have access to Gemini, you can use the basic plan subscription of the News portal cuz they also cover video analysis, which includes Gemini models.

[00:08:31]But the setup here is a bit annoying because of how the video analyze tool it routes through the auxiliary vision client, which auto detects your provider. If you pass a Google Gemini model string, the auto chain falls back to your main provider instead of routing to News, even if News is where your Gemini credentials live. So basically, if you are a subscriber of the News portal, what you need to do is to set provider News in your auxiliary.vision config cuz otherwise the tool will try your default provider again.

[00:09:07]Even for Kimi, it didn't work. So set this one up and you can use the video analyze tool. If you're using OpenRouter or another provider, make sure your config.yaml has the auxiliary.vision.provider block to set to carry those video capable models. So I'm going to give it a shot. You can do long videos as well. Actually, let's do the tier list one.

[00:09:30]We go to TUI and tell it use the video analyze tool for this video.

[00:09:42]Okay, now for some of you guys, you might run into some problems trying out the video analyze tool and it's actually a good opportunity here to show you how you might, uh, come across this problem and how you can solve it. So, when I told, uh, my Hermes agent to use the video analyze tool for that YouTube video, for our YouTube video, is that it didn't have a dedicated video analyze tool. So, it's using an alternative method to, uh, analyze that video, uh, using web extract summary by reading the transcript of my video, which is not what we want, right? It get the job done, but it you can see it's not very detailed. So, I said, uh, explicitly use the new video analyze tool from the v0.13 update. Initially, it said it didn't have a video analyze tool, uh, but thankfully it checks very carefully its skill, uh, as well as the official documentation, um, to see how to enable this. So, uh, the video tool set is enabled. However, tool changes require a new session to take effect because they're read once at startup to preserve prompt caching. Let me verify the tool is is available and try to use it. So, new session came in and it misunderstood with the vision analyze tool, but thankfully midway, uh, it got back on track and found the video analyze tool, uh, checking the actual tool implementation to how to use it properly. Uh, it exists, but it's an async function that's part of the, uh, Hermes agent internal tool registry. Let me check if it's, uh, exposed as a callable tool in my current session and also look at the test file to understand the, uh, expected interface. So, I found that it it exists in the Hermes code base, but it's not registered in the tool registry for external use. Uh, looking at the skill doc, I found the known issue. The video analyze tool has a credential routing bug, which falls back to the main provider instead of routing to the correct video capable provider. So, this is the issue that I was talking about here, right? Even though we have the news portal uh subscription, it should just, you know, use that and then give me the video use the video analyze tool to analyze my video. Uh but it it is a Hermie's issue, so it's not your fault.

[00:12:02]It understands, okay? So, it it needs a direct approach using Python to call the tool properly. Uh routing through the news provider with Gemini as documented in the skill. So, thankfully, they got it correct, uh but they noted that this is the exact credential routing bug that that was documented in the Hermie's agent skill, where the tool is falling back to the wrong provider. So, that was the fix I'm talking about. Uh explicitly routing through the news provider with the correct uh API configuration. So, it was able to use the video analyze tool uh routing through the Gemini 3 flash preview that was included in the basic plan subscription of news provider.

[00:12:40]Uh and as you can see, the the analysis of my video way more comprehensive. So, this is this is what I'm looking for. So, I think a good habit here is to tell your agent, "Can you save the correct steps so we don't run into this issue again whenever you use the video analyze tool?"

[00:13:05]Okay? All right, moving on. So, this is a very helpful durability feature, where your conversations now survive things that used to kill them. And this is very important if you're running long tasks, such as uh a Kanban worker processing a queue or a cron job with multi-step logic uh or even a research session that could span hours.

[00:13:31]Throughout that process, you don't want a gateway hiccup or a config reload to wipe your progress. That was a big issue back then. Now this is being addressed, so gateway restart is auto resume {slash} update restart.

[00:13:48]It is preserved. Source file reload is now thread targets and routing survive.

[00:13:54]Assistant metadata is being kept when branching sessions and pending update prompts are still there after restart.

[00:14:03]Okay, next up we got security hardening and don't skip this one, okay? I know that they are boring, but this specific update saw eight PO security closures that you need to update right now. So if you're running Hermes in production and if you have it connected to public messaging platforms or if you share debug logs, you need to make the update for this alone. So we're not going to go through all eight, but that these two here are the most critical one, especially if you're using Discord, okay? So redaction is on now by default.

[00:14:36]So before, secret in logs were opt in redacted. Now they're hidden by default.

[00:14:42]So if you've ever accidentally paste an API key into a chat and you've seen it in the logs, you know why this matters.

[00:14:50]And I'm talking about if you go to your dashboard and you go to your logs and you see your API keys being exposed here, yeah, you need to fix this. You need to update this. Second is Discord got a CVSS 8.1 fix. So this is a cross guild DM bypass.

[00:15:08]Basically, a malicious actor could potentially send DMs through your Hermes Discord bot across server boundaries, which is very dangerous because if you revealed your Discord token, okay, and then they DM your your Hermes agent, they can do a lot of damage there. So, be very careful. And also a helpful tip here is you can go to terminal and run the command Hermes debug share to redact at uploads. This before debug dumps could leak secrets when they're being shared, okay? So, we're near the end here. Some notable features to be aware of the cron. So, cron no agent modes. This is like a script only watch dog. So, this is good for health checks, disk space monitors, or heartbeat pings, okay? So, that's pretty much it for this video.

[00:15:59]If you find it helpful, smash up the like button, subscribe to the channel, let us know any questions you have regarding the new Hermes agent update.

[00:16:07]Subscribe to our channel, please.

[00:16:09]We really love it. All right, take good care of yourselves, ladies and gentlemen. My name is Ron. Shawn and out.