@Misbahtwts@prathamdby Context is probably too big. Grok bot handles it's context differently.
A quick fix without changing anything in the system is to have your agent write a handover with state, decisions, and next steps. Then hand it over to a clean instance
@hunvreus its even worse.
"I didn't run the routine I was supposed to".
So it knows it should have done it.
It knows it didn't.
It still doesn't
That is a level of logic I can't fathom
@nickzilllla Hermes harness is really good as a personal assistant type. Probably a bit too bloated for what you are trying to accomplish.
Did you give @pidotdev a try? its very minimalistic and a great start to create your own unique solution
@georgehanu ChatGPT work and Cursor can both receive and listen to webhooks.
I didn't try it with chatGPT and cursor, yet. But the capability is there.
So my assumption is: should work
non-verbal communication? How would that work?
LLMs are trained on words, written words.
> non-verbal - adjective
> "In a form other than written or spoken words, such as > gestures, facial expressions or body language."
Non-verbal would be hand gestures? Music notes? Gifs of me waving hand gestures? How would they even interpret that properly?
As for the MCP server side.
Your agent would initiate communication, MCP server advertises what it does and can.
ChatGPT can also subscribe to an MCP servers events.
Either way, it needs to initialise the conversation and subscribe so you can "non verbally talk with it"
https://t.co/8iZYs2rX5v
I'm currently mainly using ChatGPT, OpenWebUi, Grok Bot and Hermes.
What I like best and why ... I'm going to tell you at the end.
I mostly use ChatGPT for one-shots. Product research, concept vetting, web searches, stuff like that.
Things where I need a fast, accurate response.
ChatGPT really shines when you need one of the following:
- crawler-level knowledge of websites
- pretty pictures in my answers, stats, guides, without extra waiting.
- a confident model that does the thinking for you
- quick responses even with high reasoning levels
- refined user workflows built by OpenAI
I don't like to use ChatGPT to create plans for more complex projects, though.
It often requires you to prove your competence when giving directions. When you don't do that, it assumes incompetence and silently overwrites key decisions.
It's a pattern I notice more often in the latest models.
Then, whenever I want more privacy, I go with OpenWebUi.
It's hosted on my own servers, using Ollama Cloud - no prompt logging
I mostly use it with the latest DeepSeek and GLM flash models.
Currently, that is GLM-5.3-flash and DeepSeek v4.1-flash:1017
I use it as a replacement for search, questions, looking up private things, basic research.
Things like that.
It's usually a bit slower and way less polished than ChatGPT. But I like my privacy and it's good enough for quick queries from my phone.
By the way, send this to your ChatGPT, Claude or whatever you're using most often:
"Based on everything you know about me, roast the hell out of me. Don't hold back. Use English slang and swear like you mean it."
(prompt from @MiaAI_lab)
It's hilarious. I laughed so hard.
And after I've had enough fun, my thoughts wandered to how much ChatGPT already knows about me.
I have used it on/off since the early ChatGPT days. A month here and there over the years. And it still accurately roasted the f*ck out of me.
-- but back to topic
ChatGPT for less private and quick things, OpenWebUi for slightly slower private matters.
Then I have a personal AI assistant. Hermes Agent. It's a harness built by @nous_research.
I started using Hermes in May this year. And it has been an amazing and interesting time.
I've built and refined its skills, features, workflows and changed how it works a million times.
And every time I update it, many of my custom features and fixes become unnecessary. That’s because Hermes now has them by default.
It shows you that the Hermes devs use their own software, which is a great sign.
Hermes is my workhorse.
I use it for everything.
- Building services and scripts I need
- creating workflows and architecture
- maintaining my servers, monitoring, and systems knowledge base
- keeping my websites up to date and secure
- Researching topics
- and so much more
I use it privately, for projects, for work. It's what helps me manage everything.
I even had to give my Hermes agent a name: Lent.
I had to start naming them after I built a way for my different agents to message each other.
I'm also using Grok Bot.
I would love to tell you how much better Grok Bot works than Hermes. After all, it isn't cheap.
Sadly, my experiences have been pretty disappointing so far.
I could hardly use it at all. Support ignored my issues for weeks. Then, when they eventually fixed it, they wiped the local machine Grok Bot uses. Wiping the little work I had done.
I called my main Grok Bot "Groob". Then I built the message protocol to have it communicate with Lent, my Hermes agent.
I tried using it as a reviewing agent for research and plans. Have it do its own research and compare it with Lent's results.
Have it send interesting X posts into my morning briefing.
It's decent at agent work. I would argue my Hermes agent is still delivering better results though.
To be fair though, I've spent months refining my Hermes, while only having spent three weeks with Grok Bot. And I've had two of these weeks with limited access. So take this with a small pit of salt.
Grok Bot is decently faster, though.
Where Grok really shines is when you need access to X
- find interesting posts
- manage bookmarks
- find accounts worth following
- finding posts to answer (add them to bookmarks. Don't let your agent answer posts --- it's rude)
I'm sure I'm missing more.
I only got full access to grok-bot this week.
Aside from that, I do like the image generation from Grok.
Might currently be the best out there, since you don't have to take a dedicated course to get great results.
I also have tried other harnesses and models. DeepSeek Harness, Pi Harness, OpenCode, and other stuff.
Now, which harness and model do I like most? And why?
Hermes with the latest DeepSeek and GLM flash models.
Without a doubt, this is the best combo for me, especially compared to what else is currently out there.
Hermes is not the fastest worker and not the best coder, but it shines everywhere else.
To improve coding performance, you can even have it steer a Pi harness for you. Pi is a lean harness perfect for customisation. It's also better at coding than Hermes and can be shaped into exactly what you need, with no extra clutter.
What's your favourite harness?
And have you tried Hermes yet?
Deterministic workflows.
Do I give my agents admin rights? Hell no
I give them scoped, clearly defines permission to perform certain tasks which would need sudo.
They can only do that one thing with it, and it's for one case. Even if they turn crazy, they can only do what I gave them permission to do.
Also, deterministic vs probabilistic.
AI is Probabilistic. Ask the same question, get different answers.
Code is deterministic. Same input => same output
Build as much as possible in deterministic scripts, then have your agent use them whenever necessary
That way we go from:
What needs done -> probabilistic approach -> "Schrödingers solution"
to
What needs done -> deterministic script -> correct solution
You don't have to quit to start owning the design again.
Coding with AI and vibe coding are different things.
Vibe coding produces AI slop.
Slop is, when you let it go (the AI) .. and have it do things on it's own.
Coding with the help of AI is giving your AI clear tasks, with clear outcomes.
Try this:
- Always have it write a plan for implementation
- READ IT FULLY
- don't skim it
- tell it what is wrong and how to approach it properly
This already boosts quality a crazy amount.
Remember: LLMs are not even remotely human. They're guessing words.
The more critical context you provide,
and the better you picture the correct outcome,
the better the results will be.
I'm advocating for businesses to go local or use cloud hosted open models.
The frontier models change without notice.
Open models don't.
They release new versions, and you can choose to use older models, still.
Another benefit is to keep control over your data.
With customer data, that is a lot more important depending on location.
Get started with how to write good prompts.
Key word: Prompt engineering.
And don't listen to the people saying it's not necessary. These are the same people that learned how to interact with Ai by learning prompt engineering. Just know that old explanations might be overly complex sometimes. Ai was a lot "dumber" a while ago.
Then after experimenting with that, check out a harnesses.
Claude Code, @pidotdev, Hermes (@NousResearch), @DeepSeekHarness
My suggestion: Set up a Hermes on a local machine or VPS, connect it to telegram/whatsapp, and have it explain to you how to use it best.
Now you have an agent, living on a server, talking with you over messenger, being able to create their own automations, host back ends, dashboards, etc.
And the best? It grows better the more you use it.