How I Figured Out What to Actually Use My AI Agent For

I installed OpenClaw in the middle of all the hype. Stared at it for an hour. Never went back to it.

Sound familiar? You’re not alone. When you ask “What do you use your AI agent for?” on any forum or timeline, the answers are always some variation of “anything,” “everything,” or just “coding.” Those answers are technically true and completely useless if you’re trying to figure out your own setup.

Here’s how I actually got started.

The Audit That Changed Everything

My personal philosophy with AI is simple: I treat it like an assistant, not a replacement. It does grunt work, points me in the right direction, and I verify before proceeding. For automated tasks, I only have AI execute things I already understand how to do. That’s the line I don’t cross.

So what things do you actually use a personal AI agent for? That’s the question nobody answers well.

What worked for me was brutally honest inventory. I wrote down everything I did for a full day. Then I looked at that list. Then I expanded it over the course of a week. I asked myself:

• What took a lot of time?
• What did I have to do that didn’t add much value to my workflow?
• What are the recurring irritations I just… deal with?

That last question was the breakthrough. I didn’t just mean tech problems. I mean the softer stuff. The things that impact your life as a human being. Things I forget to do. Things I have to deal with that just make life harder.

Make that list. Then start playing with your agent.

Building the Crew

Here’s the thing about Hermes — you don’t have to run one agent. You can configure different profiles, each using a different provider and model, and switch between them instantly from the TUI or over Telegram. So I didn’t start with one agent. I started with a crew.

Tech Research Agent — I give this agent a topic and ask for a research brief with citations. Citations are non-negotiable for me because I want to go read the actual papers and source material. For example, I used this agent to learn how to do model quantization. I didn’t have it do it for me — I had it teach me how to do it myself. That agent runs on the Nous Portal with MiniMax M2.7, though I’ve also tested it with models from NVIDIA NIM.

Tech Task Master Agent — This is my anything agent. Building skills for Hermes, TUI customizations, the things that need actual execution. Currently running GPT 5.5 via my ChatGPT Plus subscription (not the API), which has been surprisingly flawless. I’ll probably keep this setup and find a backup for when I eventually hit quota limits.

Lifestyle Agent — This one surprised me the most. I live with a chronic health condition, a variation of MCAS with severe food allergies. I use this agent to scour the internet for studies and news related to this, and honestly for the most mundane thing imaginable: figuring out what to make for dinner. I cook every single meal I eat myself. Some days I’m just staring at the fridge like “not again, I have to cook.” So I’ll give it a list of recipes and have it pick one, or give it ingredients I want to use and have it suggest what to make.

Here’s the wild part: this agent runs on a local model. On an 8GB RTX 4070 in a gaming laptop. Hermes connects to it over wireless. I’m running Qwen 3.5 9B quant with a 64K context window. You’d think you’d need something bigger for something this useful, but honestly? This setup works better than I expected. I’ve also run the same model on my M1 MacBook Pro with 16GB of RAM and it holds up fine. Everyone should try running something local — you probably have more capable hardware sitting around than you realize.

The Cost Side: Doing This Cheap

I’ll be honest — I’m on a mission to do all of this as cheaply as possible. I’ve seen too many horror stories of people connecting to the Anthropic API and spending hundreds of dollars a day. No thank you.

OpenRouter — I put $10 in credits mostly to get 1,000 requests per day and 20 requests per minute on free models. The totally free tier gives you 50 requests per day which goes very fast. My current free model of choice is nvidia/nemotron-3-super-120b-a12b:free. The price is right.

Nous Portal — $10/month for their API subscription. It includes tool calling, which matters. I’ve been using MiniMax M2.7 on it. It works well and I use it sparingly since it’s metered.

Local Models — My setup is modest but effective. NVIDIA RTX 4070 with 8GB VRAM, llama.cpp serving with 64K context. Qwen 3.5 9B is my current favorite but I like experimenting with other distilled and abliterated models. LMStudio is the easiest way to get started if you want to try local models, and connecting it to Hermes is straightforward now.

ChatGPT Plus — $20/month. Connecting via my subscription and using GPT 5.5 has been almost flawless. I only did this a few days ago and I’m wondering why I waited so long.

NVIDIA NIM — If you head to build.nvidia.com/models you’ll find a bunch of free models with API access. Great for getting exposure to different models and seeing what they “feel” like before committing.

DeepSeek v4 — Everyone on my timeline has been telling me to try the DeepSeek API. The pricing is apparently incredible right now with a 75% discount through the end of May. Haven’t tried it yet but it’s on the list.

A lot of this is knowing what model will do what you need, and that comes from trial and error. Many providers are subsidizing costs or offering free access on new model releases. I’ve built skills that let me ask my agents what the current free/cheap options are on Nous Portal or OpenRouter whenever I want to experiment.

The Biggest Mistake I See

People start with the tech instead of the problem.

You don’t need a stack of RTX 3090s to get started. You don’t need to know every framework. You don’t need to have it all figured out on day one. The agents, the models, the providers — that’s all infrastructure. It doesn’t matter until you know what problem you’re trying to solve.

Go do the audit first. Figure out what actually takes up your time and what you keep putting off. That’s where your agent setup starts. Everything else is just plumbing.