Hermes Agent at home: the assistant that decides on its own how to talk to me and build my site
🔗 Learn more about me, my work and how to stay in touch: maeste.it: personal bio, projects and social links.
A week dedicated to agents, from various angles. In the deep dive I tell you about my experience with Hermes Agent, which for a few weeks now has been acting as my personal assistant on a dedicated machine at home: how I configured it, how I worked on security, and two concrete episodes in which it took initiatives that genuinely impressed me and that show what it means to have an agent with real access to your own data. In the links section you’ll find echoes of the same theme from different perspectives: antirez’s repository for local inference of DeepSeek V4, Anthropic’s new research on model explainability, the Dream mode of Claude Managed Agents that brings a form of self-improving in the cloud, the rumors about Orbit (Anthropic’s upcoming proactive assistant), Subquadratic’s claim about a 12-million-token context window (to be taken with a pinch of salt) and nine principles for writing skills that are actually useful. Enjoy the read.
My agenda
Episode 51 of Risorse Artificiali is out, with the story of Hermes Agent running locally and deciding on its own to render an HTML into pictures.
The thesis running through the episode: AGI is no longer just the model, but the integrated system (model + optimized inference + harness). Episode
On Wednesday a new interview comes out: Domenico Gagliardi (Founder and COO Kortix) explains why no software is defensible anymore in the AI era, and where value still lies (infra + data).
By now you all know our GitHub repository with tools and configurations for AI coding from the terminal on Linux. It now has its own site with single-script install at Lince.sh
We released AntiVocale (Google Play, GitHub), a software to translate voice messages into text
Solo:
On Tuesday evening I’ll be in Milan listening to Alessio talk about Local AI at the AI Meetup #14
The video of the talk I gave with Alessio at VoxxedDay Zurich has been published
On May 30th I’ll have the honor of being one of the PyCon Italia speakers
On June 12th I’ll be in Catania as a speaker at Coderful
On June 24th I’ll be in Milan as a speaker at AIConf
A month with Hermes Agent: setup, security and autonomous initiatives
For a few weeks now I have been running Hermes Agent on a dedicated machine at home, and I want to share with you my experience with an agent that constantly acts as my assistant. Those who have been reading me for a while know that in the past I had tried OpenClaw on a virtual machine, but I had left it because it was still immature and because at the time I had little time to spend on it to figure out how to make the most of it. Hermes Agent is essentially an alternative to OpenClaw: completely open source, not tied to any big tech, and with great attention to security, since the developers of the system come from the blockchain world.
For the setup I took an old PC I had at home and installed a completely fresh Ubuntu on it, dedicated only to Hermes, so that it would have no direct contact with my main machine, even though it sits on my network and has access to some of my cloud resources. I had thought about putting it on a Raspberry, but since I had this free machine available I preferred to give it a bit more resources. As a model I tried both GLM-5.1 and GLM-5-Turbo, and right now I’m stable on 5-Turbo with the coding plan, while TTS and STT run locally on the same machine, so that voice never leaves the house.
I gave it access to my personal accounts (not the enterprise ones) both on Google Workspace and on my GitHub projects, but with great care: you know that the sandboxing topic matters a lot to me. Fine-grained tokens to let it see only what I want it to see, and above all a manual hardening of the Hermes skills to remove at the script level all the operations I considered dangerous, like deleting or sending emails autonomously. I didn’t simply instruct it not to do those things: I actually removed them from the available scripts, because an agent cannot do what it doesn’t know how to do.
That said, the agent does a lot of things for me: it manages the smart devices in the house, monitors the podcast performance, checks the mail and prepares reply drafts (after the hardening above), keeps calendar and todo list with integrated reminders, manages my llm-wiki at night, does PR reviews on LINCE, summarizes articles and papers building schemas and tables and reading them aloud, and has started curating my information feeds. While I was waiting for my son at basketball practice, I asked it to rebuild my personal site, and the result is what you can find now at maeste.it.
The most fascinating thing about these agents is that, when they have a decent amount of information about your conversations or about the data you’ve given them access to, they start to take interesting initiatives. On the site, it not only proposed a style very much in line with my taste, but it went and retrieved information I had not explicitly given it. On the previous site, for example, the list of conferences I had spoken at was missing many. It had access to this newsletter, it extracted the mentions of past conferences and went on to reconstruct the complete list, including citations to slides and related videos. It really impressed me.
Another episode. While I was outside walking, I asked it to dig into an article for me and summarize it. What it decided to do on its own was a summary in Italian (because it judged that with low attention my native language was more convenient, even though we usually talk in English), and it sent it to me as a voice message on Telegram, generated by the local TTS. But for some more visual concepts it judged that voice alone wasn’t enough, and it built an HTML on the fly to represent them at their best. At that point it realized on its own that an HTML would be inconvenient to consult inside Telegram, especially on a smartphone, and so it took screenshots adapted to the format of a phone screen, sending me those instead of the whole document. Really impressive.
If you are software engineers with a sensibility for security, it’s a tool you can install with a bit of care and have fun with. If instead you don’t feel strong on that front, Hermes Agent is probably still too much in geeky territory. But if you are a geek, then it’s your territory.
Links that caught my attention this week
ds4.c
Here comes Salvatore Sanfilippo’s repository for DeepSeek 4 inference on Apple machines. I had already talked about it in past weeks, mentioning that Salvatore was working on this, and he has finally started publishing the code. The thing itself is interesting, the code itself is interesting, just like the increasingly strong trend of local inference.
Natural Language Autoencoders
Interesting research from Anthropic, which as always is attentive to explainability and in this case tries to give a natural-language explanation of the internal activations of a model.
Claude Managed Agents: Dreaming, Outcomes and Multiagent Orchestration
At Anthropic’s latest conference one of the interesting announcements was Dream mode, which lets cloud agents have a form of self-improving by reflecting on how they have been used and which skills should be modified or prioritized.
Anthropic is working on Orbit
This is little more than a rumor, but it’s said that Anthropic is working on this Orbit, a cloud mode that resembles agents like OpenClaw and Hermes Agent. Above you’ll find my concrete experience with a personal agent of this kind. It seems Anthropic too is interested in this market.
Subquadratic: a 12-million-token context window
This is exactly the article I had Hermes Agent summarize for me in an early phase, the one I was telling you about above. The article itself is interesting and talks about an evolved multi-level attention mode that allows the context window to grow up to 12 million tokens. Experience however teaches us that these things should be taken with a pinch of salt, because in the past too there was talk of research reaching 100 million tokens that didn’t go anywhere. What raises the most suspicion in the article and in the paper is that single tests or little more were run for each inference head: really too few to shout miracle.
9 Principles That Separate Useful Agent Skills From the Rest
A very interesting article that summarizes well what skills really are and when they are truly useful and interesting. If you’re writing systems that rely on skills as a standard, agentskills.io or simply Claude’s skills which are essentially the same thing, this is an absolutely must-read article.