← Back to search

NEW Hermes Delegation Update is INSANE!

AI News Today | Julian Goldie Podcast · 2026-10-28 · 9 min
relevance 100 1979 words Episode page ↗ Audio ↗
Show full episode description
NEW Hermes Delegation Update is INSANE!
✨ Episode Outline — click any point to jump to it in the episode
Problem solved
Hermes sub-agents all ran on one expensive model, so cheap lookups cost as much as hard debugging; the Advanced Delegation plugin routes each sub-agent to the right model or profile.
Benefits
  • Route simple tasks to cheap/fast models, hard tasks to Opus-class models
  • Keep sensitive client data on a local model via an OpenAI-compatible endpoint
  • Agent only sees route names and descriptions, never model IDs or API keys
  • Hermes refuses routes not on your allow list before the sub-agent starts
  • Every result reports route, provider, model and reasoning effort used
Use cases
  • SEO workflow: Search Console pulls and page summaries on Fast route, outline/page writing on Deep route, client data on Local route
  • Building a dashboard panel: researcher profile studies the data source, coder profile writes the panel and runs tests, main agent manages
  • Recommended rollout: start with Fast and Deep routes, Fast as default, review which route each result took for a week
KPIs / results
  • Requires Hermes 0.22.5 or newer (0.22–0.22.4 won't install the plugin)
  • Three delegation modes: default, model routes, profile routes
  • Setup claimed in about five minutes
Tools / build
0:00 / 0:00
Hermes Agent just got a new plugin called Advanced Delegation and fixes one of the most expensive problems with AI agents. Right now, when your agent hands work to a sub-agent, that sub-agent usually runs on one model. So for example, like a quick lookup, a file summary, a quick writing job, usually they'll go to the same big expensive model. With Advanced Delegation, the main agent can pick a different model or a whole different Hermes profile for every single sub-agent. And cheap and fast for the simple stuff, it could be slow and careful for the hard stuff, private and local for anything sensitive. And in this video, I'll show you what it does, how to set up in about five minutes, and how I'd use it inside a real agentic setup. Let's get into it. So here's what happened. News research released an official plugin for Hermes Agent called Advanced Delegation. That's two new modes to delegate task tools. That's a tool that Hermes uses when it splits a job up and hands pieces to sub-agents. Before this, there was just one mode, the default. Every sub-agent ran on the same delegation model or on the main agent's own model. Now you get three modes. Default, which is how it always worked, model routes, and profile routes. You pick one at a time. So let me break down each one. First, model routes. This is the one that most people use. So you make a short list of named routes. Each route has a name, a one-line description, and a model behind it. News gives a really nice example in the docs. So for example, it could be a route called Fast. The description says cheap and quick. And this is for lookups, for summaries, for simple edits. So it can run something like Gemini Flash through OpenRouter in the background. And then you can have a route like Deep. So this is slow and careful for hard debugging and design work. And that can run called Opus with high reasoning efforts. And then you've got a route called Local. So this is private and runs on your own machine. And that one points out a local model on your computer. Then every time the main agent hands off a task, it reads those descriptions and picks the right route for that job. The quick lookup goes to Fast. The tricky bug goes to Deep. The private client data stays local. And your most expensive model stops doing the boring work that a cheaper, smaller model could do. And here's a detail I really like. The agent only ever sees the route names and your one-line description. So it's not going to see like the model IDs, the endpoints or the API keys. And it can only pick from your list. So if it tries to pick a route that doesn't exist, Hermes actually refuses it before the sub-agent even starts. So the agent can't make up a route. And it can't wander off onto a model you didn't approve. You can also set a default route. So any task where the agent doesn't pick a route goes to, say, Fast. And if you don't set a default, those tasks just run on the main agent's own model. So every result also tells you which route it took. So the route name, the provider, the model, the reasoning effort, all of it. So you can see exactly where your work went. And it also works with OpenRouter. So OpenRouter is literally the provider in Noose's own example. And you can also use a point of route, for example, for any OpenAI-compatible endpoint, which is how the local route actually works. Now, if you want Hermes agent setup with model routes like this on your own computer, check out the AI Profit Boardroom. Here's what you get inside my four advanced delegation setup from this video. Step-by-step, the fast, deep, and local routes with the exact descriptions that give Hermes so it picks the right ones. All my best Hermes agent training in one place, from your first install to sub-agents profiles and plugins like this one. Our agensic operating system where you can plug in Hermes. Our 30-day Hermes roadmap. All our best tutorials on this stuff for live coaching week, where you can actually bring your Hermes agent and ask questions. And you can direct message me if you need help with Hermes as well. Link in the comment description or go to the aiprofitboardroom.com to get access. So back to Hermes, the second mode is profile routes. And this one is really powerful. In Hermes, a profile is a whole separate setup. So its own model, its own credentials, its own skills, and its own soul.md file, which is basically that agent's personality instructions. With profile routes, the main agent can send a task to one of your profiles. So you might have a coder profile that writes and refactors code and runs tests. You might have a researcher profile that does web research and long write-ups. The main agent can hand the coding job to the coder and the research job to the researcher. Each one runs exactly as if you'd started that profile for your task. So instead of one agent trying to be good at everything, you get a team of specialists and a manager who knows who to call. And now one more important safety note from the docs. The list of profiles is also an allow list. The agent can only ever send work to profiles you've listed. But when you list a profile, you're letting your main agent's sub-agents use that profile's credentials, its soul.md, and its terminal access. So only profiles you're happy for your main agent to actually use. It's just good sense. All right. So how do you actually set this up? So first, you need Hermes 0.22.5 or newer. That's the first version that supports this kind of plugin. Anything from 0.22 to 0.22.4 is not going to work. Hermes won't install it. So update first, then install the plugin. In the terminal, it's Hermes plugins install advanced hyphen delegation. Or in the desktop app, just go to settings, then model, then sub-agents and hit install plugin. One thing to note, plugins install per profile. So install it in the profile that does delegating, the main one. Next, run Hermes delegation config. This opens a simple menu and you can pick your mode. So default, model routes or profile routes. And if you pick model routes, you get a route editor. You can hit add a route, give it a name like fast and write a one-line description because that's what the agent reads to decide. Pick a provider you've already got credentials for, pick the model. And if you want, you can actually save the reason and effort and a speed tier. So Hermes checks each route before it saves it too. So if the provider isn't known or the effort level isn't real, it tells you right there. You can add your deep route, your local route, pick a default and you're done. And if you pick profile routes, it's even easier because it's just a checklist of your other profiles and you can tick the ones you want, add a line about each one and pick a default. And that's basically it. And if you ever want to see your setup, just run Hermes delegation config hyphen show. It prints your current mode and all your current routes. Now there's also a slash command right inside the chat. So if you type forward slash delegation, you will see your current mode and a picker to switch. And here's the clever part. When you switch modes, the change applies from your next session. When you type forward slash new. Why? Because changing the tools in the middle of a chat can break the prompt cache. And that means the next message re-reads the whole conversation and uses more tokens. So Hermes basically saves you tokens. Now, if you want it right now, you can just add dash dash now and it applies to the chat you're in. And on Telegram, Discord and Slack, you even get buttons to pick the mode. Now, how would I actually use it? So inside my authentic operating system, Hermes runs a lot of jobs, right? It does my SEO. When I ask Hermes to research a keyword and build our page, it splits it into pieces. It pulls the search console data, checks what's already ranking, summarizes top pages, writes the outline and writes a page, and then checks the page of mistakes. Now, most of those steps are pretty simple. So pulling data and summarizing pages is perfect for the fast route. Planning the outline and writing the page is where the deep route actually becomes useful. Anything with client data in it can stay on the local route on my own machine. So it's the same job, but each piece runs on the model that actually fits perfectly to it. And here's another one. Let's say, for example, I'm building out a new panel for my agentic operating dashboard. Well, the researcher profile can look up how the data source works. The coder profile could write the panel and run the test. The main agent could just manage your team. And let me put it side by side for you. The old way is every subagent runs on one model. So you either pay top price for every tiny lookup or you're cheap and your hard jobs come back half done. Or you write your own messy routing code to glue it all together. But the new way with this new update is you write a short list of routes, each with one line, saying what it's for. The main agent picks one for the right task. Simple work goes cheap and fast. Hard work goes to your best model. Private work stays on your machine. Every result tells you which model did it. And the agent can't pick anything that you didn't approve. Now, should you turn this on today? Here's my advice. Start with model routes and just two routes, fast and deep. Write really clear descriptions because the descriptions are how the agent decides. Set fast as your default. And then run your normal jobs for a week and just look at which route each result took. If deep is doing simple work, tie to learn descriptions. Once that's worked, add a local route or try profile routes as well. And the people will learn to run a whole team of agents, each on the right model, are obviously going to get way more done for way less tokens. So here's your choice. You can set this all up by hand. You can figure out the routes, write the descriptions, test yourself, or you can join the AI Profit Boardroom and have Hermes routing your subagents this week. Inside, you get the fast, deep local routes that I've talked about today with the exact descriptions set up step by step. You get the coder and researcher profile set up, ready to copy. You get the agentic operating system with Hermes plugged in so you can see every subagent's route and results in one place. You get the 30-day Hermes roadmap, daily tutorials every time news research ships an update, and four live coaching calls every week where we build your Hermes team together. Plus, you can direct message me directly when someone's online 24-7 as well. That means you can get help whenever you want. Your big model could stop doing the boring lookups today. Link's in the comment's description or go to theairprofitborn.com to get access. Thanks for watching.