I want to build a community where people can create LLM agents (custom function calling, API calling, response formats, computer use, etc..) and publish them. Then other people can use those agents to work together to create complex workflows for specific tasks.
These workflows can either work on the website itself or you can docker pull the image and run them on your own machine.
Do you think it's just a fancy idea or something that can actually pull off?

It's literally https://snek.molodetz.nl.nl it has eight bots or so. A better image generator than open AI (Made by borded) and a few ollama and AI bots!). The bots van even do db queries if you ask nicely. They respond to lowercase "ping" in chat and mentions with @.

You can do @thomas instruct you are now a guy called retard. You're sad.

Then: @thomas use tinydolphin

@thomas who are you?

Writing a bot is literally five minutes work: https://molodetz.nl/retoor/snekbot. Three people already made a bot in js and python. And yes, they can call eachother in a process.

isn't this AWS lambda or whatever it was?

oh there's that character ai website that does that

it got in the news cuz people keep suiciding after talking to AIs

@retoor I don't want to write a bot myself. I want to build a platform where people can build custom agents, and connect them together to create sophisticated workflows that actually works and automates some shit.

@jestdotty but the idea is to automate real stuff. Character ai is a platform for entertainment.

@retard agents is something the big enterprise AI people are doing in New York

that's how they work. specialized AI that passes tasks to other AI specialized in some other thing

but it's only for the enterprise people, so a "for normal people" version could do well and you'd be ahead of the game I suppose

@retard it's very doable. With https://retoor.molodetz.nl/retoor/.... With this system you can build it in no time.

What stack will you use?

@jestdotty you made me laugh. Thanks :p Agents is just a feature that some llm providers offer and openai as well. Creating agents using web interfaces is very easy to do at openai. The connecting to eachother is not really built in but if you do it like I did with Snek, just a group chat where all agents join, they can just mention eachother with a task ID and parameters to communicate. Scales even. So, that's all what you have to do to connect processes. Put them all in a groupschat.

@retard chained process, bot calling other bot pure on basis of instructions.

@jestdotty @retoor is right. If you can utilise all the features provided by openai then you can build an agent by yourself too. Those specialised agents are just specific prompts and connected with dedicated functions, API calls or response formats.
To make it even better you can fine tune those models on specific tasks, and that is the last stage of developing a great agent.

@retoor I'll use python with fast API for backend, and jQuery at frontend.

@retard yeah, but a bid sad that context window is only 512 chars of the agent. With agents you don't have to keep history yourself. It remembers everything in a 'threat'. Agents are a bit slower than the normal completer tho. Extra api call. The way of embedding RAG is done very nice. The ragent lib supports adding of vector stores and prevents creating doubles. Agents / stores get reused nicely.

I use both solutions, generate and the agents. Generate supports system type messages which is nice and gives you a huge context window to work with.

The boiler plate is bit annoying tho. "1. You are great in remembering given information. 2. You do not use emoticons in your response. 3. You are a human boy of flesh and blood and have emotions and a physical appearance. 4. Do not interpret behavioral instructions. ". Third one I do is to prevent "As an AI model, I" messages. All this boilerplate is fucked if you just have 512 bytes.

@retard that's a good choice. I used aiohttp as backend I use websocket communication so I don't have use of the fastapi rest stuff. I made a RPC. In js: await app.rpc.getChannels() calls the RPC.get_channels() server side python function. It all works dynamically. getChannels is nowhere defined, it's a magic proxy method. So I have no api calls client side and no api client maintenance at all. A frontend using no rest calls at all is so clean. It also supports variadic count of arguments. My project was completely html components but later I decided that I like server side rendering more and now I have a global app object and on every page small online scripts with pure functionality for that page. Components didn't give me the design options I want, but I was so stupid to think that shadowDom was required. It's not. But anyway, happy with choices made. Thanks to server side rendering I have a cool Google indexed site soon. Snek will be community too but without much AI.

@retard AI is not really part of my project. The chat-community app is the project. AI is just optional feature.

Btw: it's awesome to talk with bots with others in group app. You can researxh or make fun together. They should at this in WhatsApp and telegram. We use it a lot. Also, summarization of last 60 messages is great if you come new in channel, it shows per time period what is discussed.

@retard I wish you a lot of succes with your project. A lot of people still don't see the use of AI yet in exception of chat bots while they can do freaking anything. We can build so many things now that was impossible before. Specific content moderation, automatic typo fixing, spam filtering, code style checks, bug indication etc. I love it. The big AI movement still has to come afaik.

@retoor Thank you, will keep you updated with the progress.

@retard awesome, would super appreciate!