Interested in agents but clewless noob. Please help
Posted by Silver-Champion-4846@reddit | LocalLLaMA | View on Reddit | 37 comments
Hello there people. So I keep hearing about agent this, agent that, and apparently it's all the rage right now. And it also appears to be the logical next step after just chat models.
But this subreddit has been swamped with so many slop threads about "this agent is far better than anything else". Every time it turns out to be slop. Also, tools that claim to be revolutionary like OpenClaw just turn out to be heavily boosted by bots. Another factor is that I don't have a GPU, so I can't test models myself with different agents.
But I think I would really need an agent, or maybe multiple agents for different tasks I'm interested in, such as translation and assistance with novel brainstorming and co-writing. As well as a personal agent that just links all my experiences together and helps me with different random stuff.
Also let's not forget that most of the agents that are currently famous are related to coding, and I'm currently not very interested in coding agents since I haven't even learned programming myself. And I don't want to become a clueless manager of a random AI that doesn't even know how to fix the mistake that is inevitably going to arise. I actually want to know what I'm doing or what the code is doing. So I would really appreciate your assistance. Thank you.
GrungeWerX@reddit
Hey, I replied to another one of your posts above. I'm a bit busy at the moment - collaborating with my own agent on developing an AI assistant and need to look over its output - but as someone who is also not a programmer and knows very little coding - .js, .ts, python - I can say that you absolutely can take advantage of an agent for what you want to do.
The possibilities are limitless and your mind will be blown with what you can do and you'll wonder how you got along without this tech all your life. It's an amazing feeling to have someone working alongside you. It actually helps keep you on task and motivated too.
I'll post more details later, but for memory I use redis - it's the fastest and stored in RAM, which I have plenty of (96GB and an extra 32GB sitting around waiting for a bios update to make it recognizable). I also have postgres w/pgvector installed for long-term memory.
I am using n8n for orchestration. Everything's setup in Docker.
RE: Agent hype - the hype is real. Its value depends on you though. To be honest, not everyone is creative, so the only limit to an agent's usefulness and value is your imagination.
GWX
myglasstrip@reddit
Tldr, if you can't code and don't have a ton of free time, agents won't do something revolutionary for you.
The Only way the agent can interact with the world is through code. You're not going to find anything revolutionary for agents right now because if I have something revolutionary why would I give it away for free?
This means, if you're not willing to use the AI to code to build things for your agent to use then you're going to have to find someone that already is doing exactly what you find useful. Otherwise, ya, agents are useful, but at the end of the day, it's still code. If you haven't learned programming, well, I suspect you'll be blindly relying on chatgpt/etc ai to give you code and you'll learn as you go.
Either way, everything has to be tested and iterated on. I'd say, if you want the ai to actually do things amazing for you? Expect a year of work. You technically can make anything you want probably(within reason) using Ai. It's just how much time and research? The ai makes it faster you can build an app alone now. But that doesn't mean the AI just goes off and does it, that means that you're sitting at the computer 24/7 doing every single job of an organization. You never get to hand off a task and then wait for the feedback from someone else, no matter what you have to start working on the next task. Like, I made something for myself that probably would have took a person a year or two to develop before by themselves but that's still months of time by yourself gone.
You've got to look at your own free time, is there something better that you can be doing with your time than do it, but otherwise if you have the time and the patience to learn you're probably going to be in a good position learning how to make the newest tech running the world work for you. But I also recognize that this is highly Technical and not everyone's cup of tea. I have many days working on projects that I want to break everything.
Anything that you want to do as a custom project for an agent, take the initial budget of time and then multiply it by 10. You'll hit so many hurdles to overcome. Is the project actually worth it? Me personally, I won't be doing that many future ai projects, this one is specifically for the financial return I hope it delivers. Otherwise everything takes 1 month minimum to implement, you have to ask if it's worth burning that time. I have been working on a project and it's 1:00 a.m. and I say just one more bug fix and then I'm seeing the sun come up. I don't know if I can recommend that to a person to play with an AI agent just for fun unless they actually are going to get a massive reward from the project.
Silver-Champion-4846@reddit (OP)
Currently I only know some python basics. Not gonna save my axx anytime soon. Also I would need a gpu, or mix gemini free tier with free tier Openroutor somehow without getting the trouble from how different models react differently to the same agent prompt
GrungeWerX@reddit
Dude, that guy has TERRIBLE advice. I occasionally see ppl like that discourage others. There was this other guy on reddit a while back that was told by one of these programmers that what he was trying to do would take months and that he needed to learn proper programming, blah blah.
I told him to ignore that dude, and that he could absolutely make it happen and it would probably only take a day. He DM'ed me later - and updated his reddit post - that he accomplished it in 4 hours.
Hours. Not months.
I'm not a programmer. But I'm running an agent as I type this that's helping me build a personal assistant. Knowing code will of course help, but it's not a necessity and agents can absolutely be revolutionary for you.
With AI's help, I was able to build a screenwriting app that is the best app I've ever used, and I enjoy 100 times more than Final Draft, Fade In, or Trelby. If I listened to guys like this, it never would have happened.
SO, just as I've done with others, I want to encourage you to ignore the naysayers and go for it. You will absolutely find value and it will change your life.
Just remember to continue your learning. You'll reach hurdles, but you can overcome them. Listen to the do-ers, not the talkers. Ignore the naysayers. Surround yourself with those who say you can, not those who promise you can't.
Best wishes!
Silver-Champion-4846@reddit (OP)
Also, making the newest tech running the world work for me from 8gb of ram and a naive, newbie mind. What a weird thought.
LionStrange493@reddit
you're not wrong, there's a lot of hype around "agents" right now
if you're just starting, i'd skip multi-agent stuff and heavy frameworks for now
a simple setup that actually works is:
- one LLM
- a couple tools (search, file read, etc)
- a loop where it decides what to do next
that's pretty much what most "agents" are under the hood anyway
you don't need a GPU either, APIs are enough to get started
once you build something small and see how it behaves, things start to click
Silver-Champion-4846@reddit (OP)
What's a beginner-friendly (also blind-friendly) agent builder you recommend for me to try and test on? By blind-friendly I mean that I can't deal with graphics-heavy programs that build agents by drag and drop
LionStrange493@reddit
yeah for blind-friendly setups i’d skip anything visual and go code-first
simplest path is honestly just one python file:
LLM call + a couple tool functions + a loop — you can get something working in \~50 lines
if you want slightly more structure:
- openai agents sdk / anthropic tool use (very straightforward)
- smolagents from huggingface (super minimal, no UI)
i’d avoid heavier frameworks like langchain at the start, they add a lot of abstraction early
happy to share a tiny example if that helps
Silver-Champion-4846@reddit (OP)
What about memory? What do I use?
Mantikos804@reddit
Get a raspberry pizero2w. An sd card. Get on your PC talk to Gemini or ChatGPT to walk you thru using raspberry pi imager to setup pi os lite. Ssh in to pizero2w, Install Ollama. Make an account. Pull a cloud model like Gemini-3-flash-preview:cloud. Install PicoClaw, set up Ollama as provider and cloud model as the brain. Connect telegram to a bot account and talk to you “Claw” on telegram. Give it a cool name like Viper or Raptor and make an image in chatgpt for its telegram profile.
Silver-Champion-4846@reddit (OP)
Depends on whether or not it has a screen reader
Mantikos804@reddit
It can use a screen mini hdmi get the right cable. I just run them headless.
Silver-Champion-4846@reddit (OP)
Uh, screen reader, as in the program that lets blind people navigate computers.
Silver-Champion-4846@reddit (OP)
weeeeeeelll that's an adventure for the ages.
Mantikos804@reddit
Fair warning: once you do it…it’s addictive like cats. You end up having lots of ai agents and you’ll talk about what they do!!! People will think you’re nuts! But it’s fun as hell.
Silver-Champion-4846@reddit (OP)
Fare warning. I also don't have the money to pay for unlimited Gemini use.
genunix64@reddit
For what you describe, I would not start with a "full agent platform". Start with a normal chat model plus a few explicit, boring capabilities: save notes, search your notes, maybe call one or two tools. The dangerous part for a beginner is not that the agent writes code; it is that you stop knowing what state it is carrying and why it made a decision.
For writing/translation, memory matters more than autonomy. You want it to remember stable things like character names, tone rules, terminology, previous decisions, and preferences. That is different from RAG over a folder of documents: RAG is good for source material, memory is better for
Silver-Champion-4846@reddit (OP)
Better for.......... Also I also don't know of anything that has memory that I can easily use. That's what I was picturing when I said agent, or maybe half of it. Half memory, half tool calls. And I keep finding people saying vector dbs (which idk how to use anyway) are suboptimal for memory.
genunix64@reddit
I am working on plug and play mnemory system for AI agents. You might want to check https://github.com/fpytloun/mnemory It is also part of Cognis which is cloud-native agentic OS done right 🙂 https://github.com/fpytloun/cognis
Silver-Champion-4846@reddit (OP)
It seems to require an api key...
genunix64@reddit
It is opensource and self-hosted with native integrations for systems like opencode, claude-code, openwebui, openclaw, hermes, etc. It works with any OpenAI endpoint. I am personally using that with gpt-oss-120b on Groq but when you run local models you also have local openai API endpoint to use that with.
Silver-Champion-4846@reddit (OP)
I've been recommended Hermes agent. Can I just tell it to /install mnemory?
genunix64@reddit
I have hermes plugin for mnemory so it should work but never tried myself. Please if you try, let me know how it works or open Github issue and we fix it 🙂
Silver-Champion-4846@reddit (OP)
Trying to find the biggest baddest collection of free-tier apis I can dump into the model thing like Gemini free quota and Openrouter, all without somehow getting that thing where different models respond differently to the same prompt and could mess up everything
Silver-Champion-4846@reddit (OP)
I will check that out, thanks!
Ha_Deal_5079@reddit
for translation and writing id just use claude or chatgpt directly. no gpu needed and honestly most of the agent hype rn is about coding agents anyway
Velocita84@reddit
Here we go again recommending proprietary slop on the local sub...
Silver-Champion-4846@reddit (OP)
And when the ring fell, sauron screamed: NOOOOOOOOOOOOOOOO
Silver-Champion-4846@reddit (OP)
If I use gemini free api quota (which is the most generous free thingy I know of), the agent is gonna burn so many tokens and the quota with it.
SM8085@reddit
I was thinking about this yesterday since I'm trying out Hermes-Agent by NousResearch, because it has a bunch of coding skills & tools built-in. "The coders think everything is about coding. My debian friend thinks everything is about debian." It lets you deactivate those tools/skills though, although they also don't clutter it up too much either.
So far, I'm mostly using Hermes for meal planning along with Mealie (hosted on my NAS) and a Mealie-MCP (hosted on my LLM rig). Qwen3.6-35B-A3B is now capable of checking my recipes/meals and creating daily meal plans for me.
Since bots are bad at math, I originally told it to use Python to check the calorie math. It decided to write a short Python script that takes in the meals that it selects and finds combinations that fit my calorie goals. Now, the cool part is that it then decided to create a Mealie SKILL.md file that recorded things about the Mealie-MCP, and also made a resource file that included the Python it had written, for future reference.
Now, when I tell it to create a Mealie meal plan, it loads the skill it made and knows a lot of how I want things. My calorie goal. To use the Python to create meal plans that reach that calorie goal. Etc. So that's cool.
I'm not sure if the general agent Goose has that ability, to write new skills to improve itself. But you might also want to check that out for a general agent framework.
I'm probably going to load opencode into the Hermes github directory and ask it a few questions about how things work when I have questions.
Silver-Champion-4846@reddit (OP)
Goose on windows decided to crash on first launch. Typical
Stareters@reddit
Hello, sorry, because I’m also extremely new, and have very little coding experience. Local solutions to running LLMs and Agents are getting progressively better, but do not expect fast or frontier model (Claude/ChatGPT/Gemini/Etc.) thinking capabilities. A lot of recommendations for hardware stem primarily from local computing power and how much RAM or VRAM you can spare for the model’s thinking capabilities. Case in point, I shot myself in the foot by purchasing a Mac mini M4 with 16gb of RAM, hence I can only run heavily quantized models under the GGUF type which you can find online on platforms like huggingface.
For your purposes, you will hit another bottleneck, context and memory persistence. Local models offer better privacy and configuration, but have a limited context window, meaning that a model/agent’s working memory of a document, conversation, or code base in a single pass is shortened, akin to amnesia. For long term projects like translation and co-writing, you might find it to be too strained, as it could max out the context window dependent on what hardware you have for your model.
A solution that you could consider would be to have an agent running on a local model version of Gemma 4 e2b, e4b family to chip away at translating something for you, but expect it to have a minimal idea of what you are writing. Then consider tying it into a cloud LLM solution (subscriptions) for higher end thinking beneficial for brainstorming and co-writing.
Last note, if you are considering running anything locally, expect a lot of debugging as hardware incompatibilities exist. You can rely on LLMs to assist you, personally I recommend Docker’s Gordon Helper AI.
Silver-Champion-4846@reddit (OP)
Thanks for your help. I understand the high-level dynamics of local models even without testing them extensively. Re: gemma4 e2b: if I had 18 cores in my cpu, sure. But my intel i5 8350U struggles even with 2b models, around 4tokens per sec generation. And when I have 8gb ram, context is also minimal. My latitude 5590 can be upgraded to 32gb max ddr4 ram, but would that be enough?
Stareters@reddit
Upgrading to 32gb max ddr4 ram would improve context and allow you to run larger models, e2b-e4b. Token generation would still be slow, and your hardware would be prone to crashing. If you use it for purposes apart from being an LLM agent playground, it becomes difficult. You could allow it to work by turning it into a dedicated workstation that’s attached to an egpu (given you have a thunderbolt 3-4 port to connect to the dell), which would effectively move the inference from CPU to VRAM. The cost comes down to how willing you want to make your Dell work with hardware additions and upgrades. You will absolutely have to wrangle with coding, hardware issues, and general headache.
Silver-Champion-4846@reddit (OP)
Still problematic
Silver-Champion-4846@reddit (OP)
Thunderbolt? Does an old laptop like mine even have it?
Stareters@reddit
I researched a bit for this question, please take everything I say with a grain of salt. I’m likening it to the similar issues I am encountering considering I would like to use an AI agent to web scrape for a sales role. I am severely bottlenecked by RAM hence why I’m considering retooling my Mac mini M4 16GB 2024 to run Gemma 4 and Llama CPP (referring to this Reddit post ) . I’m considering a mix of local for summarizing and receiving messages, and cloud for larger more complex tasks. It would be one solution, unless I can somehow get more performance by hooking up the Mac mini to an egpu. But regardless, more costs for me. Yay.