返回列表
🧠 阿头学 · 🪞 Uota学 · 💬 讨论题

Hermes Agent 的真正用法：先找摩擦点，再配多智能体

这篇文章最有价值的判断是“个人 agent 应该从生活和工作里的具体摩擦点出发，而不是从模型、显卡和技术炫耀出发”，但它也明显高估了多智能体配置的普适性。
打开原文 ↗

2026-05-05 原文链接 ↗
阅读简报
双语对照
完整翻译
原文
讨论归档

核心观点

  • 问题先于技术 作者最站得住脚的观点是，agent 的起点不该是“我能用什么模型”，而该是“我每天哪些事又烦又不得不做”。这个判断是对的，因为多数人不是缺模型，而是缺清晰场景。
  • AI 应该当助手，不该当替身 作者明确主张“AI 负责脏活累活，人自己核实”，这个边界感比“全自动代理一切”更成熟。尤其在研究、学习、健康信息这类高风险任务里，这种保守用法明显更可靠。
  • 多模型路由比单一最强模型更实用 文中把研究、执行、生活提醒、饮食辅助拆给不同 agent，本质上是在做任务分层和成本分配。这个思路是务实的，因为不是所有任务都值得调用最贵模型。
  • 个人 agent 的高频价值常出在小场景 喝水提醒、菜谱选择、健康研究追踪这些例子听起来“不高级”，但恰恰说明真正能形成使用习惯的，往往不是宏大自动化，而是高频、低风险、即时反馈的小任务。
  • “低成本上手”成立，但被说得太轻松了 作者强调免费模型、本地小模型、订阅混搭能把成本压低，这部分有现实基础；但他淡化了配置、调试、切换 provider、维护 profile 的时间成本，这会误导新手以为门槛很低。

跟我们的关联

  • 对 ATou 意味着什么 ATou 如果在想 agent 产品或工作流，不该先卷“全能”能力，而该先拆出 2-3 个高频摩擦点场景。下一步可以直接做一版“摩擦点盘点表”，把任务分成研究、执行、提醒三类测试。
  • 对 Neta 意味着什么 Neta 如果要把 AI 变成稳定习惯，重点不是多会折腾，而是先找那些重复、低价值、但必须完成的动作。下一步可以先试一个最小场景，比如资料整理或每日提醒，而不是一上来搭复杂多 agent。
  • 对 Uota 意味着什么 Uota 会对“软摩擦”特别敏感，这篇文章最适合拿来验证“AI 是否能减轻生活心智负担”。下一步可以把 agent 用在决策疲劳最强的场景，比如饮食、计划、提醒，而不是追求技术完整性。
  • 对三者共同意味着什么 这篇文章本质上是在提供一种“任务路由”框架：高价值高风险任务用强模型，低风险高频任务用便宜模型或本地模型。下一步应该把任务按风险、频率、隐私、成本四个维度排一次，而不是凭感觉选模型。

讨论引子

1. “只自动化我本来就会做的事”这个边界是不是过于保守，还是现阶段最现实的人机协作原则？ 2. 像喝水提醒、做饭建议这种场景，算是真需求，还是被 agent 包装出来的“伪效率”？ 3. 多 agent 分工到底是长期可维护的个人系统，还是技术爱好者阶段性的折腾幻觉？

过去几周里，我一直在运行一套给 Hermes agent 用的多智能体配置。说实话，我花了点时间才走到这一步。OpenClaw 正热的时候我就装了，盯着它看了一个小时，然后再也没打开过。为什么？因为不知道该拿它做什么。那阵子我在 X 上看大家都在聊这些事，什么 Mac Mini 狂潮之类的，自己一头雾水，但又确实很想参与进去。

当你问别人，你的 AI agent 都拿来做什么，很多人会回答任何事、所有事，或者写代码。我自己也这么答过。今天想聊聊，怎么开始想清楚 Hermes Agent 能拿来做什么，这样你才能真正把它用起来。我会举一些例子，也会拿我那些有点奇怪的使用场景来说明。

怎样想清楚 Hermes Agent 能用来做什么

我自己对 AI 的看法是，把 AI 当成助手。我不用它代替思考，我用它帮我指方向。我让它去做那些脏活累活，然后自己核实，再继续往下走。至于自动化任务，我只会让 AI 自动执行那些我本来就知道怎么做的事。这是我个人的做法，你怎么用，可能会不一样。

所以，个人 AI agent 到底能拿来做什么？这就是大家都会问的问题。

我后来发现，一个特别有用的办法，是把自己一天里做过的事都记下来，然后认真看看那份清单。接着，我又在接下来一周左右的时间里不断往里面补充和展开。

我会问自己这样的问题，哪些事特别花时间，还有哪些事是不得不做，但对整个工作流并没带来多少价值。

先把这份清单列出来，然后开始拿 Hermes 去试。

那些更柔软的事

还有一个问题也值得问问自己，你日常生活里有哪些麻烦。我不是指你该在本地硬件上跑哪个模型，我说的是更柔软的那部分，是那些会影响你作为一个人的生活的事。有没有什么事你总会忘？有没有什么事你不得不处理，但就是会让生活变得更累？

我问了自己这个问题以后，又想出了几个真的很有帮助的点子。

头脑风暴完，我就开始动手了。

智能体小队

当然，我不是只从一个 agent 开始的，而是一口气弄了好几个。Hermes 有个很棒的地方，就是你可以配置不同的 profile，每个 profile 用不同的 provider 和 model，而且随时都能在 TUI 里轻松切换。自己就喜欢折腾，所以这点对我特别加分，能把不同模型并排比较，看它们各自怎么回应。

下面是我现在这支智能体小队的配置，以及它们目前使用的 provider 和 model。我平时通过 Hermes TUI 和 Telegram 来使用它们。

技术研究 Agent

通常我会把一个主题交给这个 agent，让它给我整理一份研究简报，并附上引用。对我来说，引用特别重要，因为我想自己去读学术论文和原始资料。比如我曾经用这个 agent 帮自己学习怎么做模型量化。我不是让 agent 替我做，而是让它教我怎么自己做。

这个 agent 目前跑在 Nous portal 上，用的是 MiniMax M2.7，不过以前也用过别的模型和 provider，比如 NVIDIA NIM 上的一些模型。

技术执行 Agent

我会用这个 agent 来给 Hermes 做技能，也让它昨天帮我把所有 agent 的 TUI 定制都做了。这个真的是万能型 agent。

这个 agent 目前通过我的 ChatGPT Plus Codex 订阅使用 GPT 5.5，不是走 API。之后大概率还会继续拿 GPT 5.5 来跑这个 agent，同时再准备一个备用方案，以防额度哪天用完。

顺带说一句技术类 Agent。之前有一段时间，我会混着用它们，主要是为了测试不同模型和 provider。不过接下来应该会继续按现在这个方式来，一个偏研究，一个偏执行。

生活 Agent

冒着被人吐槽的风险，我有一个 agent，专门负责在一天里的固定时间提醒我喝水。离谱吗？是。改变生活吗？绝对是。写这段的时候突然觉得，还可以让它顺便提醒我检查坐姿，再起来活动一下。这六个月我一直在试着修复这些年长期伏在电脑前把身体搞坏的问题。它会通过 Telegram 给我发提醒。

这个 agent 跑在 OpenRouter 上，用的是一个免费模型，NVIDIA Nemotron 3 Super。

生活 / 研究 Agent

我有慢性健康问题，是 MCAS 变体加重度食物过敏这一类情况。我会用这个 agent 去网上搜相关研究和新闻，也会拿它处理一些简单但烦人的事，比如，唉，今晚到底做什么吃，毕竟自己吃的每一顿饭都是自己做。像给它一串菜谱，让它从里面挑一个，或者给它一些我想拿来做饭的食材，让它帮我出点主意。有些日子真的会觉得，怎么又要做晚饭了。

说出来你可能不信，这个 agent 跑的是本地模型，挂在我那张 8GB RTX 4070 显卡上，主机是一台随手拿来用的游戏本。Hermes agent 通过无线网络去连它。老实说，这个 agent 反而是最让我惊讶的一个，因为它跑的只是个小型本地模型。我现在用的是一个 Qwen 3.5 9B quant，带 64k context。

Provider 和 Model

我现在有个个人目标，就是尽量把这套东西做得越便宜越好，甚至可能有点过头了。见过太多人直接连上 Anthropic API，然后一天花掉几百美元的那种恐怖故事。这个真不想碰。

下面是我现在在用的一些模型和 provider。

关于 Provider、Model 和成本的一点说明

个人 agent 这件事里，很大一部分在于知道哪个模型适合做什么。这件事多少有点个人化，也有主观成分，通常还是得靠试错。很多 provider 都在补贴成本，或者在模型刚发布时免费开放。很多免费模型确实会有一点代价，比如速度，但免费就是免费。

我还做了一些技能，所以如果我想再研究一下现在哪些便宜、哪些免费，直接让 agent 给我看 Nous Portal 或 OpenRouter 的当前价格就行。

Open Router - 免费模型

我往里面充了 10 美元积分，虽然其实没怎么用，但这样在免费模型上就能拿到每天 1,000 次请求和每分钟 20 次请求。完全免费的账户只有每天 50 次请求，这个额度会消耗得非常快。我现在在这里最常用的免费模型是 nvidia/nemotron-3-super-120b-a12b:free

Nous Portal - 每月 10 美元订阅

我办了 Nous Portal 每月 10 美元的订阅，主要是为了试试，结果用下来还不错。因为它是基于 API 的订阅，所以用得比较克制，不过它也支持 tool calling。现在我用的是 MiniMax M2.7。

本地模型

我的本地设备不算多强，但效果意外地还挺好。我用一台带 NVIDIA RTX 4070 的笔记本，显存 8GB，用 llama.cpp 提供服务，context 开到 64k。现在最喜欢的小模型是一个 Qwen 3.5 9B quant，不过也会去试一些别的蒸馏模型和去审查版模型。

说真的，这套配置用起来非常顺手。我也在自己的 M1 MacBook 上跑过同一个模型，那台机器有 16GB RAM。你会惊讶于手头现有设备其实能在本地跑多少东西。每个人都该试试。最容易上手的方式是 LMStudio，而且现在你也能很方便地从 Hermes 连接过去。

ChatGPT Plus 订阅 - 每月 20 美元

通过我的订阅接入，再用 gpt5.5，效果非常好，目前也没碰到任何额度问题。我也就一两天前才这么接，心里一直在想，自己怎么拖了这么久才开始用。这个几乎没什么毛病。

NVIDIA NIM – 免费模型

如果你去 https://build.nvidia.com/models 看看，会发现里面有相当多的模型都是免费的。注册一个账户，就能拿到 API key。这是一个很好的方式，能接触一大批模型，感受一下它们各自是什么风格。

通过 DeepSeek API 使用 DeepSeek v4

这个我还没试，不过 Twitter 上不少人都跟我说该去试试 DeepSeek v4 API。现在价格非常夸张，到五月底前还有 75% 折扣。还记得前面提到过补贴吧。

开始使用 Hermes Agent

如果你刚开始接触 AI agent，还在想到底该拿它做什么，希望这篇文章能给你一些头脑风暴的思路，也让你看到，其实不用花很多钱，也能很快很轻松地开始上手。

我看到大家在用 agent 时最常犯的错，就是先从技术出发，而不是先从问题出发。你不需要先搞一堆 3090 才能开始，当然，如果你真能搞到，那也挺好。

先做起来。

从你的生活开始。从你的工作流开始。从那些让你卡住的摩擦点开始。然后围着这些东西去搭 agent。

事情到了这里，才会真正变得有用。

I have been running a multi agent setup for Hermes agent for the last several weeks. Honestly? It took me a while to get here. I installed OpenClaw in the midst of the hype, stared at it for an hour, and never went back to it. Why? I didn’t know what to use it for. I watched the timeline on X about all these things going on and the Mac Mini crazy and sat there scratching my head, but I really wanted to get in on it.

过去几周里，我一直在运行一套给 Hermes agent 用的多智能体配置。说实话，我花了点时间才走到这一步。OpenClaw 正热的时候我就装了，盯着它看了一个小时，然后再也没打开过。为什么？因为不知道该拿它做什么。那阵子我在 X 上看大家都在聊这些事，什么 Mac Mini 狂潮之类的，自己一头雾水，但又确实很想参与进去。

When you ask “What do you use your AI agent for?”, many people out there will give answers like “anything” or “everything” or “coding”, and I have been known to give those answers too. Today, I want to walk through some ideas for how you get started with what to use Hermes Agent for, so you can make the most of it. I’ll give you some examples on how you can start using it for yourself, going into some of my weird use cases as examples.

当你问别人，你的 AI agent 都拿来做什么，很多人会回答任何事、所有事，或者写代码。我自己也这么答过。今天想聊聊，怎么开始想清楚 Hermes Agent 能拿来做什么，这样你才能真正把它用起来。我会举一些例子，也会拿我那些有点奇怪的使用场景来说明。

How to Figure Out What to Use Hermes Agent For

怎样想清楚 Hermes Agent 能用来做什么

My personal philosophy when it comes to AI is I treat AI as my assistant. I don’t use it to replace my thinking, I use it to point me in the right direction. I make it do grunt work, then I verify and proceed. For automated tasks, I only have AI automate and execute things I already understand how to do. This is my personal philosophy and your milage may vary based on the way you do things.

我自己对 AI 的看法是，把 AI 当成助手。我不用它代替思考，我用它帮我指方向。我让它去做那些脏活累活，然后自己核实，再继续往下走。至于自动化任务，我只会让 AI 自动执行那些我本来就知道怎么做的事。这是我个人的做法，你怎么用，可能会不一样。

So what things do you use a personal AI agent for? That’s the question we all ask.

所以，个人 AI agent 到底能拿来做什么？这就是大家都会问的问题。

What I have found to be extremely helpful was writing down things I did for a day, then taking a good look at that list. After, I went further and added and expanded to the list over the course of a week or so.

我后来发现，一个特别有用的办法，是把自己一天里做过的事都记下来，然后认真看看那份清单。接着，我又在接下来一周左右的时间里不断往里面补充和展开。

I asked myself things like “What are the things that took a lot of time?” and “What are things that I have to do, but didn’t provide a lot of value to my workflow?”.

我会问自己这样的问题，哪些事特别花时间，还有哪些事是不得不做，但对整个工作流并没带来多少价值。

Make that list, and start playing with Hermes.

先把这份清单列出来，然后开始拿 Hermes 去试。

The Softer Stuff

那些更柔软的事

Here’s another thing to ask yourself, “What are some issues in your life day to day?”. I don’t mean figuring out which model to run on your local hardware, I mean the softer stuff, the stuff that impacts your life as a human being. Are there things you forget to do? Are there things that you have to deal with that just make your life harder to deal with?

还有一个问题也值得问问自己，你日常生活里有哪些麻烦。我不是指你该在本地硬件上跑哪个模型，我说的是更柔软的那部分，是那些会影响你作为一个人的生活的事。有没有什么事你总会忘？有没有什么事你不得不处理，但就是会让生活变得更累？

I ended up coming up with a couple of more really helpful ideas when I asked this question.

我问了自己这个问题以后，又想出了几个真的很有帮助的点子。

After some brainstorming, I got to work.

头脑风暴完，我就开始动手了。

The Agent Crew

智能体小队

Of course, I didn’t just start with one agent, I started with a bunch. One great thing about Hermes is that you can configure different profiles, each using a different provider/model, and change that model easily from the TUI at any time. I like to tinker so this was a huge plus for me, to compare how models responded side by side.

当然，我不是只从一个 agent 开始的，而是一口气弄了好几个。Hermes 有个很棒的地方，就是你可以配置不同的 profile，每个 profile 用不同的 provider 和 model，而且随时都能在 TUI 里轻松切换。自己就喜欢折腾，所以这点对我特别加分，能把不同模型并排比较，看它们各自怎么回应。

Here is a break down of my current agent crew, and what provider/model they use currently. I access them using the Hermes TUI and over Telegram.

下面是我现在这支智能体小队的配置，以及它们目前使用的 provider 和 model。我平时通过 Hermes TUI 和 Telegram 来使用它们。

Tech Research Agent

技术研究 Agent

Generally I use this agent a topic and ask for a research brief, along with citations. For me, citations are key, because I want to go read the academic papers / source material. For example, I used this agent to help me learn how to do model quantizations, I didn’t have to agent do it for me, but I had it teach me how to do it myself.

通常我会把一个主题交给这个 agent，让它给我整理一份研究简报，并附上引用。对我来说，引用特别重要，因为我想自己去读学术论文和原始资料。比如我曾经用这个 agent 帮自己学习怎么做模型量化。我不是让 agent 替我做，而是让它教我怎么自己做。

This agent uses on the Nous portal currently, with MiniMax M2.7, but I’ve had it use other models/providers in the past such as models off of NVIDIA NIM.

这个 agent 目前跑在 Nous portal 上，用的是 MiniMax M2.7，不过以前也用过别的模型和 provider，比如 NVIDIA NIM 上的一些模型。

Tech Task Master Agent

技术执行 Agent

I use this agent for building skills for Hermes, and I also had it do all my TUI customizations for all of my agents yesterday. This is really the anything agent.

我会用这个 agent 来给 Hermes 做技能，也让它昨天帮我把所有 agent 的 TUI 定制都做了。这个真的是万能型 agent。

This agent currently uses GPT 5.5 via my ChatGPT Plus Codex SUBSCRIPTION, not the API. I will most likely keep using GPT 5.5 for this agent, and come up with a backup for if/when I run out of quota.

这个 agent 目前通过我的 ChatGPT Plus Codex 订阅使用 GPT 5.5，不是走 API。之后大概率还会继续拿 GPT 5.5 来跑这个 agent，同时再准备一个备用方案，以防额度哪天用完。

A note on Tech Agents: I was using them interchangeably at one point for lots of model/provider testing, but I think I’ll be continuing to use them this way going forward, one as the researcher, and one as the executor so to speak.

顺带说一句技术类 Agent。之前有一段时间，我会混着用它们，主要是为了测试不同模型和 provider。不过接下来应该会继续按现在这个方式来，一个偏研究，一个偏执行。

Lifestyle Agent

生活 Agent

At the risk of being roasted, I have an agent that’s job is to remind me to drink water at certain points throughout the day. Ridiculous? Yes. Game changing? Absolutely. As I’m writing this I think I’m going to have it also prompt me to check my posture (I’ve spent the last six months trying to fix what I’ve wrecked spending years hunched over a computer) and take movement breaks. It sends me messages on Telegram to remind me.

冒着被人吐槽的风险，我有一个 agent，专门负责在一天里的固定时间提醒我喝水。离谱吗？是。改变生活吗？绝对是。写这段的时候突然觉得，还可以让它顺便提醒我检查坐姿，再起来活动一下。这六个月我一直在试着修复这些年长期伏在电脑前把身体搞坏的问题。它会通过 Telegram 给我发提醒。

This agent runs off of OpenRouter using a free model – NVIDIA Nemotron 3 Super.

这个 agent 跑在 OpenRouter 上，用的是一个免费模型，NVIDIA Nemotron 3 Super。

Lifestyle / Research Agent

生活 / 研究 Agent

I’m stuck with a chronic health condition, a variation of MCAS / severe food allergies. I use this agent to scour the internet for studies and news related to this, as well as for simple stuff like ugh what should I make for dinner tonight since I cook every meal I eat myself. Something simple like giving it a list of recipes, and having it respond with one, or giving it a list of things I want to cook with and having it give me ideas on what to do with it, because some days I’m just like oh no not again I have to make dinner.

我有慢性健康问题，是 MCAS 变体加重度食物过敏这一类情况。我会用这个 agent 去网上搜相关研究和新闻，也会拿它处理一些简单但烦人的事，比如，唉，今晚到底做什么吃，毕竟自己吃的每一顿饭都是自己做。像给它一串菜谱，让它从里面挑一个，或者给它一些我想拿来做饭的食材，让它帮我出点主意。有些日子真的会觉得，怎么又要做晚饭了。

This agent runs off of a local model believe it or not, on my 8GB RTX 4070 card, hosted in a random gaming laptop. Hermes agent gets to it over the wireless network. I would say I’ve been the most “impressed” with this agent since I’m running it on a small local model. I’m using a Qwen 3.5 9B quant with 64k context.

说出来你可能不信，这个 agent 跑的是本地模型，挂在我那张 8GB RTX 4070 显卡上，主机是一台随手拿来用的游戏本。Hermes agent 通过无线网络去连它。老实说，这个 agent 反而是最让我惊讶的一个，因为它跑的只是个小型本地模型。我现在用的是一个 Qwen 3.5 9B quant，带 64k context。

The Providers / Models

Provider 和 Model

I am on a personal mission to do this as cheap as possible, to the point where I may be shooting myself in the foot. I’ve seen too many what I call horror stories of people just connecting to the Anthropic API and spending hundreds of dollars a day. No thank you.

我现在有个个人目标，就是尽量把这套东西做得越便宜越好，甚至可能有点过头了。见过太多人直接连上 Anthropic API，然后一天花掉几百美元的那种恐怖故事。这个真不想碰。

Here are some of the models and providers I currently use.

下面是我现在在用的一些模型和 provider。

A Note On Providers/Models/Cost

关于 Provider、Model 和成本的一点说明

A lot of personal agents AI is knowing what model will do what you need it too, and that is somewhat personal and subjective, some trial and error is often required. Many providers are subsidizing costs, or providing models for free upon release. Many of the free models do have a bit of a trade off such as speed, but hey, the price is right.

个人 agent 这件事里，很大一部分在于知道哪个模型适合做什么。这件事多少有点个人化，也有主观成分，通常还是得靠试错。很多 provider 都在补贴成本，或者在模型刚发布时免费开放。很多免费模型确实会有一点代价，比如速度，但免费就是免费。

I have also built skills and simply ask my agents to show me current pricing on Nous Portal / OpenRouter if I want to play around more with what is cheap or free currently.

我还做了一些技能，所以如果我想再研究一下现在哪些便宜、哪些免费，直接让 agent 给我看 Nous Portal 或 OpenRouter 的当前价格就行。

Open Router - Free Models

Open Router - 免费模型

I added 10 dollars in credits that I don’t actually use to get 1,000 requests per day and 20 requests per minute on free models, the totally free account gets you 50 requests per day which goes very very fast. My free model of choice here is currently nvidia/nemotron-3-super-120b-a12b:free

我往里面充了 10 美元积分，虽然其实没怎么用，但这样在免费模型上就能拿到每天 1,000 次请求和每分钟 20 次请求。完全免费的账户只有每天 50 次请求，这个额度会消耗得非常快。我现在在这里最常用的免费模型是 nvidia/nemotron-3-super-120b-a12b:free

Nous Portal - 10 Dollar a Month Sub

Nous Portal - 每月 10 美元订阅

I got the 10 dollar a month Nous Portal subscription to experiment with, and it has been working well. Since it is an API based subscription I use it pretty sparingly, but it does also include tool calling. Right now I’m using MiniMax M2.7.

我办了 Nous Portal 每月 10 美元的订阅，主要是为了试试，结果用下来还不错。因为它是基于 API 的订阅，所以用得比较克制，不过它也支持 tool calling。现在我用的是 MiniMax M2.7。

Local Models

本地模型

My local equipment is meh, but it works surprisingly well. I use a laptop with a NVIDIA RTX 4070 that has 8GB of VRAM, and llama.cpp to serve with 64k context. Right now my favorite small model is a Qwen 3.5 9B quant, but I like to experiment with other random distilled and abliterated models too.

我的本地设备不算多强，但效果意外地还挺好。我用一台带 NVIDIA RTX 4070 的笔记本，显存 8GB，用 llama.cpp 提供服务，context 开到 64k。现在最喜欢的小模型是一个 Qwen 3.5 9B quant，不过也会去试一些别的蒸馏模型和去审查版模型。

Honestly, this setup works very well, I have also run this same model on my M1 MacBook with 16GB of RAM. You would be surprised at what you can run locally with what you already have. Everyone should try it. LMStudio is the easiest way to get started, and guess what, you can connect to it easily from Hermes now.

说真的，这套配置用起来非常顺手。我也在自己的 M1 MacBook 上跑过同一个模型，那台机器有 16GB RAM。你会惊讶于手头现有设备其实能在本地跑多少东西。每个人都该试试。最容易上手的方式是 LMStudio，而且现在你也能很方便地从 Hermes 连接过去。

ChatGPT Plus subscription - 20 a month

ChatGPT Plus 订阅 - 每月 20 美元

Connecting via my subscription and using gpt5.5 is working very well, and I have not run into any quota issues. I only did this a day or two ago, and I’m wondering why I waited this long. This is almost flawless.

通过我的订阅接入，再用 gpt5.5，效果非常好，目前也没碰到任何额度问题。我也就一两天前才这么接，心里一直在想，自己怎么拖了这么久才开始用。这个几乎没什么毛病。

NVIDIA NIM – Free Models

NVIDIA NIM – 免费模型

If you head over to https://build.nvidia.com/models, you will see that quite a number of models are free. Sign up for an account and you can get an API key. This is a great way to get exposure to a bunch of modes and see what they “feel” like.

如果你去 https://build.nvidia.com/models 看看，会发现里面有相当多的模型都是免费的。注册一个账户，就能拿到 API key。这是一个很好的方式，能接触一大批模型，感受一下它们各自是什么风格。

DeepSeek v4 via DeepSeek API

通过 DeepSeek API 使用 DeepSeek v4

I haven’t tried this yet, but numerous people on twitter have told me to try the DeepSeek v4 API, the pricing is incredible right now at a 75% discount through the end of May (remember I mentioned subsidies?)

这个我还没试，不过 Twitter 上不少人都跟我说该去试试 DeepSeek v4 API。现在价格非常夸张，到五月底前还有 75% 折扣。还记得前面提到过补贴吧。

Getting Started With Hermes Agent

开始使用 Hermes Agent

If you are stating and wondering what to do with an AI agent, I hope this article gave you some brain storming ideas, and showed you that you can get started pretty quickly and easily without breaking the bank.

如果你刚开始接触 AI agent，还在想到底该拿它做什么，希望这篇文章能给你一些头脑风暴的思路，也让你看到，其实不用花很多钱，也能很快很轻松地开始上手。

The biggest mistake I see people make with agents is starting with the tech instead of the problem. You don’t need to get a stack of 3090s to get started (but hey if you can grab some go for it).

我看到大家在用 agent 时最常犯的错，就是先从技术出发，而不是先从问题出发。你不需要先搞一堆 3090 才能开始，当然，如果你真能搞到，那也挺好。

Just start doing.

先做起来。

Start with your life. Your workflow. Your friction points. Then build agents around that.

从你的生活开始。从你的工作流开始。从那些让你卡住的摩擦点开始。然后围着这些东西去搭 agent。

That’s where this actually becomes useful.

事情到了这里，才会真正变得有用。

I have been running a multi agent setup for Hermes agent for the last several weeks. Honestly? It took me a while to get here. I installed OpenClaw in the midst of the hype, stared at it for an hour, and never went back to it. Why? I didn’t know what to use it for. I watched the timeline on X about all these things going on and the Mac Mini crazy and sat there scratching my head, but I really wanted to get in on it.

When you ask “What do you use your AI agent for?”, many people out there will give answers like “anything” or “everything” or “coding”, and I have been known to give those answers too. Today, I want to walk through some ideas for how you get started with what to use Hermes Agent for, so you can make the most of it. I’ll give you some examples on how you can start using it for yourself, going into some of my weird use cases as examples.

How to Figure Out What to Use Hermes Agent For

My personal philosophy when it comes to AI is I treat AI as my assistant. I don’t use it to replace my thinking, I use it to point me in the right direction. I make it do grunt work, then I verify and proceed. For automated tasks, I only have AI automate and execute things I already understand how to do. This is my personal philosophy and your milage may vary based on the way you do things.

So what things do you use a personal AI agent for? That’s the question we all ask.

What I have found to be extremely helpful was writing down things I did for a day, then taking a good look at that list. After, I went further and added and expanded to the list over the course of a week or so.

I asked myself things like “What are the things that took a lot of time?” and “What are things that I have to do, but didn’t provide a lot of value to my workflow?”.

Make that list, and start playing with Hermes.

The Softer Stuff

Here’s another thing to ask yourself, “What are some issues in your life day to day?”. I don’t mean figuring out which model to run on your local hardware, I mean the softer stuff, the stuff that impacts your life as a human being. Are there things you forget to do? Are there things that you have to deal with that just make your life harder to deal with?

I ended up coming up with a couple of more really helpful ideas when I asked this question.

After some brainstorming, I got to work.

The Agent Crew

Of course, I didn’t just start with one agent, I started with a bunch. One great thing about Hermes is that you can configure different profiles, each using a different provider/model, and change that model easily from the TUI at any time. I like to tinker so this was a huge plus for me, to compare how models responded side by side.

Here is a break down of my current agent crew, and what provider/model they use currently. I access them using the Hermes TUI and over Telegram.

Tech Research Agent

Generally I use this agent a topic and ask for a research brief, along with citations. For me, citations are key, because I want to go read the academic papers / source material. For example, I used this agent to help me learn how to do model quantizations, I didn’t have to agent do it for me, but I had it teach me how to do it myself.

This agent uses on the Nous portal currently, with MiniMax M2.7, but I’ve had it use other models/providers in the past such as models off of NVIDIA NIM.

Tech Task Master Agent

I use this agent for building skills for Hermes, and I also had it do all my TUI customizations for all of my agents yesterday. This is really the anything agent.

This agent currently uses GPT 5.5 via my ChatGPT Plus Codex SUBSCRIPTION, not the API. I will most likely keep using GPT 5.5 for this agent, and come up with a backup for if/when I run out of quota.

A note on Tech Agents: I was using them interchangeably at one point for lots of model/provider testing, but I think I’ll be continuing to use them this way going forward, one as the researcher, and one as the executor so to speak.

Lifestyle Agent

At the risk of being roasted, I have an agent that’s job is to remind me to drink water at certain points throughout the day. Ridiculous? Yes. Game changing? Absolutely. As I’m writing this I think I’m going to have it also prompt me to check my posture (I’ve spent the last six months trying to fix what I’ve wrecked spending years hunched over a computer) and take movement breaks. It sends me messages on Telegram to remind me.

This agent runs off of OpenRouter using a free model – NVIDIA Nemotron 3 Super.

Lifestyle / Research Agent

I’m stuck with a chronic health condition, a variation of MCAS / severe food allergies. I use this agent to scour the internet for studies and news related to this, as well as for simple stuff like ugh what should I make for dinner tonight since I cook every meal I eat myself. Something simple like giving it a list of recipes, and having it respond with one, or giving it a list of things I want to cook with and having it give me ideas on what to do with it, because some days I’m just like oh no not again I have to make dinner.

This agent runs off of a local model believe it or not, on my 8GB RTX 4070 card, hosted in a random gaming laptop. Hermes agent gets to it over the wireless network. I would say I’ve been the most “impressed” with this agent since I’m running it on a small local model. I’m using a Qwen 3.5 9B quant with 64k context.

The Providers / Models

I am on a personal mission to do this as cheap as possible, to the point where I may be shooting myself in the foot. I’ve seen too many what I call horror stories of people just connecting to the Anthropic API and spending hundreds of dollars a day. No thank you.

Here are some of the models and providers I currently use.

A Note On Providers/Models/Cost

A lot of personal agents AI is knowing what model will do what you need it too, and that is somewhat personal and subjective, some trial and error is often required. Many providers are subsidizing costs, or providing models for free upon release. Many of the free models do have a bit of a trade off such as speed, but hey, the price is right.

I have also built skills and simply ask my agents to show me current pricing on Nous Portal / OpenRouter if I want to play around more with what is cheap or free currently.

Open Router - Free Models

I added 10 dollars in credits that I don’t actually use to get 1,000 requests per day and 20 requests per minute on free models, the totally free account gets you 50 requests per day which goes very very fast. My free model of choice here is currently nvidia/nemotron-3-super-120b-a12b:free

Nous Portal - 10 Dollar a Month Sub

I got the 10 dollar a month Nous Portal subscription to experiment with, and it has been working well. Since it is an API based subscription I use it pretty sparingly, but it does also include tool calling. Right now I’m using MiniMax M2.7.

Local Models

My local equipment is meh, but it works surprisingly well. I use a laptop with a NVIDIA RTX 4070 that has 8GB of VRAM, and llama.cpp to serve with 64k context. Right now my favorite small model is a Qwen 3.5 9B quant, but I like to experiment with other random distilled and abliterated models too.

Honestly, this setup works very well, I have also run this same model on my M1 MacBook with 16GB of RAM. You would be surprised at what you can run locally with what you already have. Everyone should try it. LMStudio is the easiest way to get started, and guess what, you can connect to it easily from Hermes now.

ChatGPT Plus subscription - 20 a month

Connecting via my subscription and using gpt5.5 is working very well, and I have not run into any quota issues. I only did this a day or two ago, and I’m wondering why I waited this long. This is almost flawless.

NVIDIA NIM – Free Models

If you head over to https://build.nvidia.com/models, you will see that quite a number of models are free. Sign up for an account and you can get an API key. This is a great way to get exposure to a bunch of modes and see what they “feel” like.

DeepSeek v4 via DeepSeek API

I haven’t tried this yet, but numerous people on twitter have told me to try the DeepSeek v4 API, the pricing is incredible right now at a 75% discount through the end of May (remember I mentioned subsidies?)

Getting Started With Hermes Agent

If you are stating and wondering what to do with an AI agent, I hope this article gave you some brain storming ideas, and showed you that you can get started pretty quickly and easily without breaking the bank.

The biggest mistake I see people make with agents is starting with the tech instead of the problem. You don’t need to get a stack of 3090s to get started (but hey if you can grab some go for it).

Just start doing.

Start with your life. Your workflow. Your friction points. Then build agents around that.

That’s where this actually becomes useful.

📋 讨论归档

讨论进行中…