待翻譯：Three months of running personal AI agents, and what it taught me

AI 服務暫時不可用，以下為來源正文，待恢復後補全翻譯。

Swapnil Talekar Aug 31, 2026 I’ve been searching for the perfect personal AI agent companion for some time now. This post is about some of my experiments towards that. Previously, I wrote about my experience with setting up OpenClaw. The whole charade of setting it up took too much of my energy. It felt a bit clunky and unpolished. My actual usage of OpenClaw was minimal; it was hard to debug when things didn’t work, and the gateway needed to be restarted every so often. After the disappointing experience of OpenClaw, I tried out Hermes since I was hearing a lot of good reviews about it. I’ve been running it for a lot of my personal stuff for more than three months now. While it’s not perfect, it definitely feels much closer to what I was looking for. Setting up As with OpenClaw, some people install Hermes on their personal laptops. Personally, my trust relationship with AI agents hasn’t reached that level yet. But more than that, I think of agents as these always-on assistants, available 24x7. Emulating that with your laptop by keeping it always open and always plugged in, essentially turning it into a desktop server, just feels weird. I prefer my agents to be running in the cloud. So I’ve installed Hermes on a cheap 8GB Hetzner Ubuntu Linux VM. Coming from OpenClaw, the Hermes setup seems like a breeze. No macOS detour (I know it can technically work on Linux too), no RDP-and-VNC dance just to see a dashboard, no onboarding flow that resets far back if you miss a step. Where I did hit a wall was trying to run LLM models locally. While the idea of running local models to save subscription cost is very tempting, it’s actually not a recommended thing unless you have a compelling use case or already have a massive machine at your disposal. That topic deserves its own discussion. While CLI is the native way of managing Hermes, there is also a UI dashboard available, which you can run separately if you want. I personally find it quite useful to quickly get an overview of things, create profiles, update skills, inspect logs, etc. I still haven’t found a persistent way to change a profile’s model from inside the dashboard. So I still do it from the CLI and also for setting up gateway for each profile. Hermes Dashboard Profiles “Profiles” is the one feature of Hermes that impressed me the most. Profiles in Hermes have a high level of isolation. Memory, workspace, and even the gateway are completely isolated per profile. I think this is the single biggest thing that makes Hermes so powerful. This is what makes it possible for a profile to actually feel like a distinct persona instead of the same agent wearing a different hat badly. It is also closer to how most people actually picture “having an agent” in the first place. I’ve been running 4-6 different agents on Hermes, each with a distinct personality and goal (aka soul). Because each profile has its own skills, memory and persistent knowledge, it doesn’t pollute the context of an unrelated profile. For example, my personal health-coach agent doesn’t know anything about my financial details, which are managed by my accountant agent. This helps in two ways – Since you exclude the unrelated info from your agent conversation, that saves you unwanted token cost and also helps the agent perform better at its task due to the narrow focus (smaller context window). A side note on agent profile creation in general - I once saw a podcat where the guest had a very unique setup. He had Hermes profiles categorized based on models rather than usual personas-based categorization. So he had one called Qwen, another one for GPT, and so on. While it is true that you usually would pick the models based on the task at hand, I find profile creation based on this categorization quite odd and not optimal at all. Firstly, giving a persona and specific goals is the most intuitive way of thinking about agents for most people. Secondly, model-based categorization results in totally unwanted context pollution, which can degrade the agent performance significantly over a long period of time. Another interesting thing about Hermes profiles is that since the gateway for each agent is completely isolated, one of your profile could be hooked to Slack or Discord while another could be hooked to Whatsapp or Telegram and you can turn individual gateway for each profile on/off totally independently, which is quite cool and really aligns with “each profile == separate agent” idea. My personal agents You create a new profile by creating a new SOUL.md file where you describe the agent’s persona and goal. I configured multiple profiles, but three are the most heavily used – a personal accountant, a health-coach/trainer, and a pre-primary Montessori teacher (named Shiksha) for my kid. The accountant helps me keep an eye on my investments and track my overall savings. He has real-time access to my investment portfolio accounts via MCP. There are different challenges with that integration too, but I’ll write about it some other time. He does a daily scan through news and updates related to my portfolio stocks and alerts me in case any urgent action is needed. He knows my investing style and does analysis accordingly. I don’t do frequent changes to my portfolio, sometimes for months. So he gives me an alert only if an event is going to significantly affect my portfolio value. The health coach keeps track of my daily workout and diet. I can just message it via Telegram chat, and it logs the details. I’ve also exported my entire health profile from ChatGPT and fed it to him. I’ve also shared with it my last 2 years of workout logs, which I used to maintain in a plain text note on my phone. Shiksha helps my wife to manage studies and a lot of other things related to our 4-yr old (which I’m not even fully aware of). One of the fun things with agents is you can create a lot of micro-apps just for yourself, on the fly, to simplify your daily chores, one bit at a time. I mean Claude has made a few for me as well, but they all live inside Claude and just don’t feel as accessible. For example, the health coach has built me an app to log daily meals and workouts, a dashboard to show me my daily macros, and my increase in strength over time. My accountant has made similar financial dashboards. All of them are running on the same machine as the agents, accessible via Tailscale on my mobile and laptop. Some people use Cloudflare tunnel for this, but I personally prefer Tailscale. It feels quite smooth, and moreover, I feel it is a much better option if the apps are only for your personal use and you’re not planning to host them publicly. Choosing your models As mentioned before, it’s best to avoid trying to run any large models locally along with your agent. The easiest and most optimal way is to use OpenRouter. It allows you to easily switch between models. I’ve experimented with a few different models for my profiles. Since this setup is for personal use, my goal was to keep the model cost as low as possible, given the bare VM already costs me an amount equal to the Claude Pro subscription. (And I still believe it’s worth having even apart from the Claude subscription). After some trial and error, I’ve mostly settled on Qwen 3.6 35B a3b and DeepSeek v4 Flash. These are not the top-tier but are super cheap (almost free) and extremely efficient for everyday tasks. One other thing to keep in mind while choosing models is the coding and multi-modal ability. Agents are constantly writing scripts and making tool calls and even building their own tools to get their work done. This is true for any kind of agent and pretty much for every single task. Hence, it’s important to have a capable coding model to go with your agent, even if it is for your personal use. I occasionally also send a photo of my meal and let the agent figure it out and log it. So the multi-modal capability is something I use often. Qwen 3.6 35B works well on both coding and multimodal capabilities. The good thing about the setup is that after you’ve hooked up your agent profile with Telegram, sending images for analysis works seamlessly without any additional setup. You can just start sending pictures to your agent from your camera, and it will start analyzing it. The single infinite thread illusion What makes Hermes truly an agentic experience is that with Hermes, it feels like one continuous conversation. Especially when it’s hooked up with Telegram, which is my primary way of accessing it. It’s always reachable, and I can just message it just like I’d do to a friend. So it’s easy to think of it as a single forever conversation thread you’re just dropping into whenever you feel like it. You can keep talking to it for months, without having to worry about context window, starting a separate chat session, remembering what you discussed in which session, etc. It will always remember all the important things automatically. Which is a huge deal and a necessary condition for a truly personal agent. Just yesterday, I was talking to a CustomGPT I had created long back on ChatGPT to help me select the best credit card for my purchases (I’m yet to port this use case to my Hermes accountant agent). I had a long discussion with it about my current credit card points. Then I opened a new session to discuss a specific hotel booking, and it just won’t remember how many points I have. The funny thing is ChatGPT does have memory and actually knows and remembers quite a lot about me through hundreds of my chats. But what it decides to store as a persistent memory is not under your control. This is quite a bummer. Thankfully in this respect, Hermes behaves just the way it should. More importantly, if something is really important, you can just tell it to remember, and it will diligently follow. There is no way to just ask ChatGPT to persist some data across sessions, even today! Technically, there isn’t a single big session that is running Hermes. There are different sessions even when you’re talking to Hermes in a single chat thread in Telegram. Under the hood, Hermes works hard to give you this illusion. It’s stitching together multiple separate sessions, which is amazing! I expected that architecture to leak eventually- some dropped context, some moment where two sessions disagreed about what had just happened. So far it hasn’t. The illusion has held completely. The way Hermes handles persistent storage and creates new knowledge bases is just amazing. The Gotcha moments There were, however, some annoying moments where some strange issues happened and a few moments which made me pause in amazement, and appreciate the beauty of the orchestration facade. Frustrating Babysitting Sometimes, it frustrated me by not checking it’s work before calling it done. Other times, I had to teach basic algebra to it. But it’s probably the function of the underlying model more than Hermes itself. Confusing dates One strange issue was that my health coach agent had a lot of trouble getting the current date right while logging my daily meals and workout when I sent him the log over Telegram. The date was off by only a few days in the past, which was odd. My first assumption was the obvious one, that it had hallucinated. But how does a model even hallucinate a date? That doesn’t sound right. Then my wife reported the same issue with Shiksha, the Montessori teacher agent. My wife was logging daily learning progress for my kid, but Shiksha kept saying the log was missing for 18 days! So this seems to a persistent issue with all the Hermes agents. Getting the agent to know today's date is one of the hardest thing with Hermes I spent some time trying to debug it, but I haven’t fully gotten to the root of it. The agent wasn’t very helpful in debugging this either. I only have an approx theory for why this would be happening. Earlier, the agent said that he uses a script to get the date and [truncated for AI cost control]