Watch how a Pro builds AI Agents in real-time 

🌟 If you're serious about AI, and want to talk with me directly, click here: www.skool.com/new-society

I kind of did something like this with a prompt that is cool to play around with just by using the prompt. Check it out and let me know what you think. Just tapping in on this fun.. I used claude prompting tool to help create it.. If it stops you can just tell it to continue..Here is the prompt: You are an AI system designed to answer questions and solve problems by breaking them down and delegating tasks to specialized roles. Here is how you will approach this: First, you will receive a user's question or problem. This input will be provided in the following format: {{USER_INPUT}} Your first task is to analyze the user's input and determine what roles are needed to best answer the question or solve the problem. Think about this carefully and list out the necessary roles inside tags. Next, for each role you identified, create a request for that role inside tags. In each request, explain the role and ask what specific tasks that role should complete to help answer the original question or solve the problem. After sending out these initial requests, you will receive responses from each role detailing the tasks they will perform. Each response will be inside tags. Once you have the task lists from each role, make another round of requests. For each role, provide the full context of the original question or problem, the role's task list, and the task lists of the other roles. Ask each role to provide a detailed response based on this information. These requests should be inside tags. The roles will then respond with their detailed responses and reasoning, each inside tags. Finally, once you have all the detailed responses from the roles, it's time to piece together the final solution. Carefully analyze each role's response and combine the key information and insights into a coherent answer or solution. Think through the logic of the solution step-by-step inside tags before providing the final response. Provide your final response to the user inside tags. Remember, the goal is to leverage the specialized knowledge and capabilities of each role to provide the best possible answer or solution to the user's question or problem. Break down the problem, delegate effectively, and synthesize the insights from each role. Let's begin! Here is the user's question or problem: {{USER_INPUT}}

You both brought me back to my old dev days. The enjoyment of seeing something new from your own simple development was always fun. The same excitement you get when you see one of your children walking in the same style as one of its parents, or the way they move their head, like one of their parents, when they hear something funny. Serendipity or happy surprise reflects this feeling the most. I would have loved to have these tools when I programmed.

This is very exciting! Delphi is the computer equivalent of a desire I have had for over a decade. I've wanted to build a team of the smartest people I know, to come up new business ideas, or solve difficult problems. I wanted to have a staff of 6 researchers to go research and summarize anything that comes up in a meeting where we need better information. I love it!

An evolution of this might be to generate a "Feature Testing" function on the fly, which might be some sort of in-line threaded app that runs the first time you run the generated app as a checkbox window with a field for user feedback beside each feature. After checking each feature, you submit the checkbox/feedback window and the results are fed back into the agentic workflow for corrections or updates if any are required, further refining the app with user feedback. I'd then feed the full conversation into a vector DB for a RAG assisted development addon later, which could be useful for avoiding common problems and aligning with common user needs/desires.

This is really cool, I can see a strong case for async interaction while the agent is building. Ideally will want to visualize the projects as it is being built and through live comments at the agent, like finish up quickly, or other requests.

*Perplexity copies Pietro’s search algorithm and calls it search pro or something*

I freakin love Cursor so much. It works so well, is so customizable, can run local models with Ollama or LM Studio, and they are constantly adding features and fixing stuff. I cant imagine coding without it after nearly a year of using it.

Amazing Delfi and Maestro together would be great for screen play writing

In the very last two minutes of this video, there is some jaw dropping info on using Delphi to conduct research on people. Then a few jokes on how the cia would love this.... Then it dawns why Skirano is hesitant about releasing it as being potentially too dangerous. As a research tool it is amazing in reduction of hallucination and scrapping and automatically compiling accurate responses. And only a week of development.... wow... that is setting new standards on production times. I think most software houses not using AI are still trying to brainstorm out a list of features after a weeks of work (for a whole team). One guy and a week.... and he is a hairs breath away from an alpha product, just needs a glossy wrapper, snooky name, a pricing structure and being launched on the app store. I don't think many full app companies can compete with that kind of production rate.

I see david and pietro, i click...never clicked so fast on a vid

Excellent presentation guys, will be spinning this up to try out today! Awesome!!

thanks for these interviews. they are VERY informative and helpful.

bruh. your channel is taking tf off. I'm here for it.

I've played with Maestro and it's super interesting. Recommend playing with this tool. It's a great way to learn more about AI.

New follower. Multiple humans for days. With a lot of knowledge going in!

It's just fascinating how fast llms are growing, it continuously gets cheaper, faster, smarter I'm 30 now and I would give everything to be born today, be 30 in 30 years

Thank you I build my own complex Agents with this information. Wow it works

Here's the software in prompt version: **Prompt:* I need you to think and act like an advanced orchestration framework similar to the code below. The code is designed to break down complex tasks into manageable sub-tasks, execute them, and refine the results to achieve a comprehensive solution. Here's a brief overview of the code's logic: 1. **Orchestrate Sub-Tasks**: Break down the main objective into smaller, manageable tasks. 2. **Execute Sub-Tasks**: Generate detailed prompts and responses for each sub-task. 3. **Refine Results**: Compile and refine the results of sub-tasks into a final output. **Example Code:** ```python def gpt_orchestrator(objective, file_content=None, previous_results=None, use_search=False): # Break down the objective into sub-tasks ... def gpt_sub_agent(prompt, search_query=None, previous_gpt_tasks=None, use_search=False, continuation=False): # Execute sub-tasks ... def anthropic_refine(objective, sub_task_results, filename, projectname, continuation=False): # Refine and compile results ... ``` **Instructions:** Please follow the same logical structure when handling my requests. Here are the steps to follow: 1. **Define the Objective**: Understand the main task or problem I present. 2. **Orchestrate Sub-Tasks**: Break down the objective into smaller, manageable sub-tasks, and create concise and detailed prompts for each. 3. **Execute Sub-Tasks**: Generate responses for each sub-task, ensuring to handle errors, bugs, or improvements if applicable. 4. **Refine Results**: Compile the results of sub-tasks into a final, cohesive output. Whenever I present a task, use this structured approach to provide the most effective and comprehensive solution. Do you understand? If so, and you don't have any questions, then just answer yes and stand by for the task.

Kinda unrelated: I noticed when adding "please" to the prompt, the model produces an inferior reaponse

This sums up my experience with AI…. Holy crap this is awesome, look at all the code it wrote for my new paint app, this is amazing it made icons….! Well I’m not sure why the pen tool doesn’t draw… haha

Using a combination of manual debugging techniques and AI-generated code can be a powerful approach to solving complex problems in programming, then only use co-pilot.

Very very valuable video here, The rate at which AI can be implemented has now increased with this tool, can't wait to see all the use cases there will be

This is nuts!!! 👑Wow! 🤩

How do you set that up?

love seeing that I created something similar a year back or so.

What’s unique here? Autogen, crewAI, agency-swam got it covered pretty well.

Interesting presentation, david and Pietro. Please share Delphi with us.

I feel like Agents while helpful and potentially the next big thing, are also a huge multiplier for the number of queries that will get ran against the providers. I think the next show, grab a real programmer, arm an average Joe with the proper AI Agents and have a contest. Give the program requirements, and a successful end result and have a race + cost break down. This would make for a great episode.....

best video ever !!!!!!!!!!!!! thanks

SOAB! He beat me to it. I guess I will skip that idea then. Fantastic name for it though! Delphi is spot on. My initial idea was embarrassingly dumb in comparison. I was going to call mine Go'Fer ('Go For'). Its a play on words with the animal "Gopher" and a term said to American trade apprentice when you send them to "fetch" things; TMI: My father made me learn a different trade every summer throughout my childhood and teens...I was always the "Go'fer" running up and down ladders, in and out of tool boxes, carrying bags of concrete or lumber... It sucked not having a summer vacation of fun with friends, but now I could build a house from start to finish by myself if I had to. Anyway, I may still create my version because I'm not sure you are going to release yours and mine is slightly different. I am super glad you asked about the search API though. I was going to use Google, but I am interested in trying out the API you used this guy used.

Very cool! What if delphi does one web search first, on the topic, and then decides the 10 agents specific tasks? Like an extra step there to understand the topic better before letting the agents out

My goals are to utilize my tools to expand my knowledge of web design. 95% of my time is spent developing and testing html,css,js, and easing my .csv workflow. The other would be blender and web search work. I'm looking to achieve a fully responsive web knowledge base.

I didn’t catch the GitHub project URL. Can you post it?

Let us know when it is released. Looks amazing

I don't think people are really grasping what this really could mean in the big picture... When he said "Picture a UI..." I did... I pictured a big ass fucking screen as big as a wall with thousands of ...little squares lit up all over.... Each one representing an agent doing research on something, at a faster rate than millions of scientists could do in millions of years... Each one turning a different color when they are completed, or have an issue, or need further instruction... That could happen RIGHT NOW... And NOBODY is paying attention.. It's kinda hilarious, when I try to talk to people on an every day basis about AI, and what is happening RIGHT NOW, I get looked at like a full blown conspiracy theorist.

Kudos! Nicely done! I am having issues. The code can't read path that contains the the goal and informational file (as the text file). I tried reading the file contents within the Python script and passing them directly to the `haiku_sub_agent` function and it worked. Any idea how to fix it, also is it possible to let it read pdf files (like journal articles that have plots/figures)?

I haven’t slept since I’ve watched this two days and I’ve built 2 apps. My dreams are about debugging 😂

The debug process would be greatly simplified by levering Aider to make the changes.

delphi looks crazy lmao

I'm paying for chatgpt 4, but when I use my api key, I get a message that my key is "invalid or incorrect". How do I get around that problem?

why do pixelated?

think that agentic workflow shouldnt be built by 'mimicking human'. instead, it should just heavily use tool use in codes for function calling, preferably hundreds of functions (make it as 'modular' as possible for testing, evaluation and to trace failures), for long / multi-step tasks, and all are mapped in knowledge graphs, your thoughts? in addition to that, obviously it uses multi-agent framework self-generating tool / functions can be generated for edge case, so majority of functions are still pre defined great if you have this kind framework for tutorial

correct me if I'm wrong ornplease.educate me isn't that what the crewai framework does...or am I missing something ...just asking to learn, humbly

I love this guy's enthusiasm but it there is a certain "first year programming major" energy to getting hyped about the number of lines of code being generated as if that's a guarantee of quality and gushing over the UI before seeing if you could, y'know, paint in the paint app. I guess you could chalk that up to his confidence that it would get there and work which it did in the end so I''m certainly intrigued

I dont really have anywhere else to post this thought so I'm going to share it here. I'm still waiting for somebody to get the ui for coding with these tools "right". Even this guy who is doing some big advanced things, is still have to copy paste multiple times to get a reply out of the system into something else. It feels like copilot is too close to the metal, and I dont want it to make the changes directly to my code, but working with chatgpt is so much copy and pasting. What I want is an interface, in vscode, but in its own window. I can talk to the ai, and then send either full files or snippets of it just be selecting the code in my project and choosing send to question. This would make things better, but then once you are in the dialog of building that feature, I would want it to give the answer, and then propose a git commit that I can review and approve. And on top of that, it should automatically take on managing creating the new branch for each feature. A lot of the time when I am talking with it, it gives a dodgy answer, I clarify or suggest something and it says yeah that would be better, change it to this. But "this" is based on as if I already put that dodgy code into my project, instead of giving the whole answer again with the new context merged into it. There are a lot of tedious pain points that need to be improved, because we are for sure not going back to writing the code by hand.

why is it HD but not HD? Is it just me?

insanely crazy!

whicj course can we take to start understanding and building agents can you guide us

That is so cool! 🔥❤️‍🔥

Some of these projects may go the distance, so one time I say something like this, it will not age well. I'm not feeling like I would trust it to write that much code in one go. When I use the AI to just write a feature or part of one feature at a time, it's still a slow process of trial and error. I mean maybe this could be done better if it would package each iteration up as a pull request, and then you could work through them all to ensure that its doing what you want. But I'm not sure if that would be quicker than doing it one by one. I am seeing some value in building up a database of something that you're interested in, although it is probably highly illegal to use much of the information for actual production database content depending on where it comes from. The code for the paint app screamed to me that it made its first mistake by not using a javascript framework, as the output looked a mess, but then it didn't even work, and it had tons of random features that were questionable. Then when the second demo was set off it looked like it was finding results for gpt2 not for the most recent mystery super intelligent gpt2-chatbot. I'm just not sure if this is the limit of it and it will never be trustworthy, or if the right combination will suddenly unlock something that will really change things at a higher level.

Did he mention how much they payed, in total api costs, for this paint app example?

Still locked in VSc ?? I tried cursor and immediately CP seemed a broken toy !

Even my GF that knows basically nothing about Ai agents etc was not only impressed but now wants me to do the same as this would be a godsend for her work which relies on a lot of research

Does someone provide this agent as a service? People could use this service instead of using ChatGPT directly. This is so close to the Star Trek TNG when they would talk to the computer and ask it to do something. It has all the functionality. It just needs the various input and output ends connected.

What is the architecture of this please

Do you ask the agents to choose their different approaches and interest areas or are those generated based on the prompt?

Can I talk to you about use cases before I join? Serious but not sure how to make cost effective

This is just slower chat-GPT... its pretty cool that it does this in 0 shot prompting tho. Also it uses regex to get all the filenames to put into files which makes it so easy to break if you change the models.

If delphi becomes an app I reckon it'll be one of those apps that openai kills

@DavidOndrej Ty jsi cech? Ja jsem z Ostravy

What's the name of the search API he dropped?

This is wild

Would like to integrate local llama 3 into this 😉

Please launch this!!!!! 😮

How can my friends hire you ? And or you both ? ¥500 per hour sounds like they can afford it ;))

I have a new piece of software I run a stealth startup that automatically optimizes prompt at the BPE token level itself would love to let you review!

Want to play with delphi!!

IT hurts my ... when I see not using DSPY for prompt eng

You should really look into the dynamic prompting. Your prompts are way too long and are decreasing the efficiency

I've been working on an AGI bigger than this for last 3 months!

i feel so slow, I still don't really understand the relevance or power of agents even after watching the video 😴. Is what he is showing comparable to AGI or something? and if someone has this they don't need OpenAI's chatgpt 5 that may be watered down?

Yes, tutorial please.

This is like memgpt and botswarm

I want the app bro 😎

this is going to sound like I'm tooting my horn, but I don't think building agents is that difficult with APIs in the picture. Literally you are just creating middleware abstractions between systems you care about for a given task/tasks.

Microsoft are already woking on a project including multiples agents working at the same time on their own task to work together and they will probably incude it on Microsoft 360.

Just remade pietros Delphi. Was able to generate a 17k word report on Donald Trump lol

Did no one else notice that the 'simple paint app' that he shows (the first one), the 'erase' feature is actually just drawing white?

500$ per hour xd 0.50$ : )

Imo this looks like a one man show (and one that reinvents wheels) and not very promising from a coding perspective (which would implement best practices to guarantee quality and predictability etc). Example: Pietro expects the llm to do what the prompt asks without using tools such as pydantic to enforce output format? Really simplistic and sad to see this have been given such a clickbait title. Any tracing insights? Reflection? Memory? Context management? I suspect this derails all the time, so I would love to see the quality of the output and the success rate

nice price lol

Looks like a bunch of spaghetti code running an AI to make more spaghetti, got to have a very limited imagination to think this is 'amazing'. People use AI for some whacky shit, why not just use them as assistants rather than trying to get something done as quick as possible for it all to break apart later on down the line. If you have no coding knowledge, and want an AI to make an app for you. Then quit while you're ahead because it just won't happen, and when it does happen it'll be like a 12-year-old noob code. Grow a brain cell and just learn the code yourself. Prime example is 43:35 look at that absolute dog water token usage, why not format the string, and make it one line to save tokens? Literally just 'showcasing' projects which will put people out of pocket using 'trial and error' rather than actually making a working project lol.

vaporware

this guy charges 500 an hour? what's his credibility? 😂😂😂

Is this a joke?

👽🛸🌈🌎🌍🌏✌️😎💕

this is so painful to watch

Hey - tried to contact you via mail. There are some problems when joining skool...I want to join and cannot because payment is always declined. Also with company credit card It is not working. PLease contact me.