Hermes Agent was new to me, so I built something small and real on it and paid attention to what the harness made easy.
What I built
A personal YouTube curator. Every morning a scheduled job scrapes my logged-in home feed and watch history, an agent ranks it, and I get a tiered digest on Telegram (learning, infotainment, entertainment) with a duration and a short reason per pick. A second agent maintains a small wiki of the topics and channels I keep coming back to, and my replies on Telegram steer the next morning's digest, either by changing the shape of the digest, or adding extra guidance on what my preferences are at the moment. I can also go deeper on any pick from the digest by asking for more detail from the video's transcript.
I built this to run on small, non-frontier models, so the whole point was control rather than raw intelligence. The idea is to take a modest model and keep it tightly steered on one narrow task, and that is what Hermes is good at: giving each step just enough scope to do its job and nothing more. Here is a single run:
What my project needed, and what Hermes gave me
Building blocks I did not have to write. There is an overwhelming amount of pre-built skills, and scanning them turns up useful ones. I used the youtube-content and llm-wiki skills.
A gateway and a scheduled job. Both are first-party in Hermes, so I did not have to build the plumbing for either. The cron job runs a deterministic fetch script first, and only then does the agent wake up to reason over what was collected.
A way to handle unsafe internet data. Its hooks give you a lot of granularity. I was able to give only the subagents that this particular job spins up Python write access to a single directory on my machine, and nothing else. That let me trust small local models with a pile of fresh, unfiltered YouTube data every day without worrying about prompt injection. The guard that enforces this lives in the repo if you want to see how it is wired.
A taste profile that improves over time. Hermes classifies any feedback I give as memory-worthy or not and folds it in. My preference for investigative infotainment became part of my profile, and if I want tomorrow's digest to carry less brain-off content, I just say so on Telegram and it lands.
A digest I can reshape by asking. Taste is one lever, the format of the digest is another. When I want to change the shape of the digest itself, I ask the agent in chat and it edits the skill that defines it.
The backbone, briefly
None of this is revolutionary, but the project falls into three clean, reusable pieces that I would keep for anything similar:
- A cron job can run a deterministic fetch script, so the model never does the fetching and only reasons over data that already exists.
- Subagents with isolated contexts, and very limited tool sets, are safer. The transcript one gets a terminal narrowed to a single command, the wiki one gets file access and no shell.
- Agents can then distill the fetched information with the user's help through something like an LLM wiki, so knowledge accumulates instead of being rebuilt from scratch each run.
Swap the fetch script and the topic, and the same three pieces carry over.
Some extra things I like about Hermes
There are also some smaller touches. The harness nudges the agent when it looks stuck, and it sends a token budget warning when a run goes on too long. Its own footprint is modest too: the system context it spends out of the box is around 13k tokens, which leaves a small model real room to work. I have run my local models on both Claude Code and Hermes, and while the two are built for different purposes, Hermes feels built with SLMs in mind. That has made me want to try coding-agent harnesses built for SLMs too. My experience with Claude Code and SLMs has been underwhelming, but Hermes showed me that when the scaffolding is designed around them, small models can be far more impressive.
Repo: https://github.com/pariidanDKE/HermesYoutubeCurator
If you have built on Hermes, I would like to hear how you are handling memory and untrusted input in your own agents!
Top comments (0)