Transcript
00:00:00Ever since Hermes Agent was released, people have been setting up all kinds of workflows on it.
00:00:04And to make an already powerful agent even better, people have also built a lot of skills for it.
00:00:10Some of them have even become some of the most trending projects on GitHub.
00:00:14Now you might wonder why you'd need to install more skills in the first place,
00:00:17because Hermes already comes with a lot of them and it can even create new ones from your workflows.
00:00:22But the skills Hermes creates only come from the processes it learns from you.
00:00:25These other skills are built by people who have already solved problems Hermes can't handle on its own,
00:00:31and they've packaged those solutions into skills you can use.
00:00:34Now if you're new to the channel then welcome, we're a software company and this is AI Labs.
00:00:38And in this video we're going to go over the skills that actually make using Hermes agents so much better.
00:00:43But before we start with the first one, you need to know this,
00:00:46even though we're using these skills with Hermes, most of them can be used with any AI coding agent.
00:00:51Now if you've been working with coding agents, even if it's Hermes or literally any other agent,
00:00:56they all have the same problem.
00:00:57If you're having a long conversation, your agent can forget the details you mentioned earlier,
00:01:02so you have to give it the same instruction a second time.
00:01:04And this happens more often when you're using Opus models with Hermes,
00:01:08because they tend to forget things more quickly.
00:01:10This happens because there are too many messages in your context window,
00:01:14and the model has to pay attention to all of them.
00:01:16When that happens, it can lose track of the smaller details and you get worse answers.
00:01:20To solve that problem, there is a skill called Planning with Files.
00:01:24It keeps the plan where the agent can't ignore it.
00:01:27Instead of keeping the plan in the chat and taking up space in the context window,
00:01:31this skill converts the plan into three files in your project folder.
00:01:34The first one is TaskPlan.MD, which basically splits your main task into smaller tasks
00:01:40so the agent can do them one at a time.
00:01:42Then there's Findings.MD, which holds every problem the agent faced and how it solved it.
00:01:47If the same problem happens again, the agent can use those learnings to avoid it.
00:01:51And the last one is Progress.MD, which records how far the agent has gotten.
00:01:56Now, you might think keeping your plan in files isn't new,
00:01:59and most of you might already be doing it.
00:02:01What sets this skill apart is that the agent can't ignore the plan.
00:02:04That's because it uses a hook,
00:02:06which forcefully inserts the three files into the context window,
00:02:09which is why the model can't choose to ignore it.
00:02:12Now, most agents don't support hooks, but Hermes is one of the few that does.
00:02:16And according to their own testing, the agent performed much better with the hook running
00:02:20because it didn't drift away from the task that it was assigned.
00:02:23The install commands for all the agents are listed in the repo,
00:02:26which will be listed in the description below.
00:02:28Once you have it installed, it'll be added alongside all the other skills.
00:02:32And the hooks are going to be automatically installed in your agent.
00:02:35The skill file explains what the three files are,
00:02:38where they should be, and what each one does.
00:02:40It also includes all the rules the agent needs to follow,
00:02:43so it knows what it's working with.
00:02:44So once you have this installed,
00:02:46you can run the skill and give it a prompt for what you want the agent to do.
00:02:50Here you can see that the agent started by creating the task plan,
00:02:53the findings, and then the progress files as well.
00:02:56Then it asked us questions, and we answered them as we went.
00:02:59After that long session, it finalized the plan and gave us the three files.
00:03:03But before we go into the next skill,
00:03:04it would be great if you subscribe to the channel and hit the hype button.
00:03:08This small gesture of support goes a long way for us.
00:03:11You might already know that you can use multiple coding tools with Hermes.
00:03:15For example, you can connect your codex subscription to Hermes agent and use its models.
00:03:19You might also know that Hermes already has a clawed code skill that sends tasks directly to it,
00:03:24and it has the same kind of skill for codex as well.
00:03:27But the problem with these skills is that Hermes won't use them unless you ask it to.
00:03:31You also have to say which tasks should go to each agent in your prompt.
00:03:35But there is another skill that sends a task to any agent that is already installed on your system.
00:03:40And this skill acts as an orchestrator between all of them.
00:03:43It also figures out which of your installed coding tools is best for which task and gives the work to that tool.
00:03:49This is actually a set of skills.
00:03:51The main one is called delegate setup, which coordinates the other agents.
00:03:54It also comes with dedicated delegation skills for different coding agents, including clawed code, codex, and cursor.
00:04:00These agents run in a mode where they don't ask permission before running any command,
00:04:05so the coding agent doesn't stop and wait for you to approve every change.
00:04:08To install this skill, run the install command.
00:04:11It asks you to choose from 18 skills, so you can select the tools installed on your system.
00:04:16Once installed, you will see that Hermes has the delegate setup skill ready for use.
00:04:20To use it, you just have to run the skill with the slash command, type in the prompt, and give it to Hermes agent.
00:04:26First, it identifies which coding tools are installed.
00:04:28Then it gives you a complete map of what area each agent will handle, along with the exact model name and effort level.
00:04:35Once you approve the plan, it starts running the skills.
00:04:37It then sends each task to the agent picked for it.
00:04:40After every agent reports its findings, Hermes gives you one collective report and marks the task as complete.
00:04:46Now you might already know that your Hermes agent runs terminal commands.
00:04:50Each terminal command is run, it returns an output that the agent has to read.
00:04:5490% of the time, you don't even need all of the output that command produces.
00:04:59You only need a certain part of it.
00:05:00The rest just bloats the context for no reason.
00:05:03And just like we mentioned before, having too much information in the context window makes it harder for the agent to focus.
00:05:09That's where RTK comes in.
00:05:10It cuts the output of the terminal commands the agent runs and keeps only the parts the agent needs,
00:05:16so the output gets filtered before it ever reaches the context window.
00:05:20The database of the terminal commands agents use most often, so it knows how to filter their output.
00:05:25For example, if you run tests, the agent only needs to see the failed test results and not the passing ones.
00:05:31RTK does that by default and keeps only the failures.
00:05:35But the thing is, it's not actually a skill.
00:05:37It's a separate tool you install on your device and run from the terminal.
00:05:41And it's on this list because it makes sure your agent gets only the right context.
00:05:45If you want, you can also install it directly and tell the agent to use it.
00:05:49But since agents tend to forget what we ask, we actually created a skill that tells the agent to route commands through this tool instead of running them the usual way.
00:05:57This skill also contains the details the agent needs, along with a reference for each command so it knows which commands it can use and how to use them.
00:06:05You can add this skill to your Hermes setup.
00:06:07Once you've done that, you can give Hermes any task and it will run every command through RTK instead of using the usual method.
00:06:14Only the output the agent needs reaches the context window, so Hermes can give you more accurate answers.
00:06:20But before we move on to the next one, let's have a word by our sponsor.
00:06:23Plural.
00:06:24Running production infrastructure usually comes down to one thing, the 3am alert and digging through Kubernetes logs to find what broke.
00:06:31Plural fixes that with workbenches.
00:06:34They're AI agents that actually operate your infrastructure for you and they run inside your existing access controls, so nothing happens that you haven't already allowed.
00:06:42Here's what sold us.
00:06:43We wired it into our observability and when an alert fired, the agent didn't just ping us.
00:06:48It investigated on its own, dug through the logs and metrics and came back with a real diagnosis and clear next steps.
00:06:54And Workbench hooks into the tools you already run, like Datadog and Slack, so it can kick off on its own the moment an alert fires or on a schedule you set.
00:07:02And instead of writing prompts from scratch, you hand them ready-made skills from Plural's library.
00:07:07It's basically an on-call engineer that never sleeps.
00:07:10Try Plural at the link in the description.
00:07:12Now the next skill is by Google and it's called Mantis.
00:07:15Mantis is a set of security review skills that your agents can run.
00:07:19It contains a variety of skills, each focusing on a different area of security review.
00:07:23This setup works no matter what app you're building, be it a mobile app or a web app, you can install it and use it for whatever you're working on.
00:07:30The installation command is listed in the GitHub repo.
00:07:33When you run the install command, you will be asked to choose between the many specialized skills it has.
00:07:38It's better to install all of them because they are separate pieces of the whole review process.
00:07:42Once you've installed it in your setup, you'll see all of the installed listed.
00:07:46For example, there is a skill called Mantis plan and it plans out your security review process before you actually begin.
00:07:52It also contains a skill called Mantis reflect that goes back over all the findings and makes sure the failures are documented properly.
00:07:59Now you might think if Hermes is an always running agent and we don't usually code with it, why would we even need to add Mantis there?
00:08:06But the reason is that once your app is hosted, it needs security reviews again and again and that isn't something you want to remember to do yourself.
00:08:14And that's what makes Hermes the best place to run them.
00:08:16It has cron jobs, which are basically tasks that run on their own on a schedule.
00:08:21You can set one up that runs this review set against the app you've hosted.
00:08:25And because Hermes is running all the time, it can alert you on Slack or any other connected platform if it finds a problem.
00:08:31We also have a similar setup for our own project.
00:08:33Here you can see that we gave Hermes our project in the prompt and told it to run the full set over the code.
00:08:39It keeps finding issues and applying fixes, which keeps the app healthy for us and ensures that it doesn't have any bugs.
00:08:45If you've ever tried researching with agents, you might have run into a huge problem.
00:08:49Your agent is denied access to certain content.
00:08:52This happens because a lot of sites like Reddit don't want models using their content.
00:08:56But those sites are really valuable since they have real users talking about problems and their opinions and experiences, which makes the information useful for research.
00:09:05To give agents access to that content, there is a skill called Agent Reach.
00:09:09It gives Hermes access to the sites that are usually restricted, including Reddit, GitHub, and many others.
00:09:15It uses free methods, so it doesn't cost you anything.
00:09:18It was also the number one trending project on GitHub recently.
00:09:21Now you can install this in your Hermes setup and it'll give you access to more than 15 platforms.
00:09:26It can search Instagram, Facebook, Twitter, and Reddit.
00:09:29Once you have it installed, you can use it directly from Hermes and ask it to search for any topic on any platform.
00:09:35It runs its own scripts, finds the relevant information, and then gives you the answers that you need.
00:09:41Now, as you already know, Hermes comes with more than 90 skills built in.
00:09:45And aside from that, when Hermes sees a workflow in your chat that could be useful in the future,
00:09:50it actually creates a skill out of it and then it can use that skill in future runs.
00:09:54It sounds good because it's adapting to you automatically, but there is a problem with this.
00:09:58The only part of a skill that actually goes into the context window is its name and description.
00:10:04They get sent with each of your messages so that the agent knows these skills exist.
00:10:08Now, you might say that the name and description make up a small chunk and wouldn't affect the context window at all,
00:10:14but if you have over 100 skills and you've been using Hermes for a long time,
00:10:18all those names and descriptions start consuming a huge portion of the context window.
00:10:23So, to solve this issue, there's a whole set of skills built to make the agent more productive.
00:10:28But the one we're most interested in is the skill retriever skill.
00:10:31It replaces the full list of skill prompts with only the most relevant skill descriptions,
00:10:36which causes the context size to drop.
00:10:38So, basically, it runs another step before Hermes calls the model.
00:10:42It splits your message into small chunks that a model can understand,
00:10:45does the same to every skill you have, and then matches your message to the skills list.
00:10:50The skills that score highest are the ones it sends.
00:10:52Their docs put numbers on it.
00:10:54They say that with the whole list going in, it used around 11,000 tokens.
00:10:58After this, it uses only 2.3 thousand, so they claim it saves more than 9,000 tokens every turn.
00:11:04Now, this helps you in two ways.
00:11:05First, there's a lot less going into the context window with every message you send,
00:11:09so it costs you less to run.
00:11:11And second, the agent isn't distracted by skills that have nothing to do with what you asked,
00:11:16so you get better answers.
00:11:17This comes as a plugin, and Skill Retriever is one of the skills inside it.
00:11:21Once you've installed the plugin, you can enable Skill Retriever from the plugins area.
00:11:25Then the skill list going into the context will be much shorter than it used to be.
00:11:29Now, the RTK skill that we created in this video can be found in AI Labs Pro, which is our community.
00:11:35So if you found value in what we do and want to support the channel, this is the best way to do it.
00:11:39The link's in the description.
00:11:41That brings us to the end of this video.
00:11:43If you'd like to support the channel and help us keep making videos like this,
00:11:46you can do so by using the super thanks button below.
00:11:49As always, thank you for watching, and I'll see you in the next one.