🎯 Key Takeaways for quick navigation: 00:00 *🤖 Introduzione ai metodi di creazione di un agente personalizzato* - Panoramica sui metodi sequenziale e gerarchico per sviluppare agenti personalizzati, - Vantaggi e svantaggi di ciascun metodo. 02:28 *🛠️ Strumenti e modelli per lo sviluppo di agenti* - Utilizzo del modello GPT-4 turbo e strumenti di ricerca gratuiti, - Implementazione di funzioni per salvare i contenuti e comunicare lo stato al termine delle operazioni. 04:28 *📝 Processo di ricerca e scrittura con gli agenti* - Descrizione dettagliata degli agenti di ricerca e di scrittura utilizzati nel progetto, - Interazione tra gli agenti e l'utente per definire gli argomenti di ricerca e produzione di contenuti. 06:00 *🔄 Integrazione e passaggio di informazioni tra agenti* - Uso della memoria condivisa e del callback per tracciare le operazioni degli agenti, - Definizione di limiti alle iterazioni per prevenire loop infiniti. 08:00 *🎯 Definizione dei compiti e aspettative dei risultati* - Assegnazione chiara dei compiti agli agenti e delle aspettative sui risultati, - Importanza della specificità nell'impostazione dei compiti per evitare ricerche indesiderate. 10:02 *🗣️ Interazione iniziale con l'utente e inizio del processo* - Avvio del processo con la richiesta di un argomento specifico all'utente, - Primi passi nella ricerca e nell'elaborazione dell'argomento richiesto. 12:54 *📊 Revisione e ottimizzazione del processo basato sugli agenti* - Analisi dei passaggi effettuati e del feedback ricevuto per migliorare il processo, - Considerazioni su possibili miglioramenti e modifiche al flusso di lavoro degli agenti. 15:18 *🔄 Cambio al modello gerarchico e problemi correlati* - Introduzione del modello gerarchico e dei problemi emergenti nel chiedere temi all'utente, - Adattamento del processo per incorporare più agenti e affrontare questioni di ambiguità. 18:22 *🛠️ Risoluzione dei problemi e iterazioni finali* - Risoluzione di problemi nel salvataggio e nella comunicazione dei risultati, - Riesame finale del processo e delle prestazioni degli agenti. 20:50 *🚀 Conclusioni e considerazioni future sul progetto* - Valutazione finale del progetto, considerazioni sulle prestazioni e miglioramenti futuri, - Suggerimenti per migliorare la qualità e l'efficacia dei modelli utilizzati. Made with HARPA AI
Hi thanks for the walkthrough. Would you share the total cost of these crews? Also, I have a feeling that this research project may not fully reveal the collaboration capability of the agents. The final output is seemingly a brief summary of the Jamba paper. Did you get to do a comparison with the summarization result using zero-shot GPT-4?
cost for running them quite a few times was $2.3 i didn't work out the per run cost. My guess is around 1/10 of that or less. The difference here in regards to the paper is it did its own search and found the blog post and used it. I din't give it anything more than that. This was also a somewhat simplified example of what I would normally do too, to show the basics etc. In regards to cost Haiku is way better for that.
Hey Sam, loving the vids. Have you used the feature where you pass task outputs into other tasks via the task context variable? Does this make much of a difference? This can be seen on the crew ai website.
hi, thanks for your video. I'm wondering if you know how to make crewai's output long and detailed? Maybe dozens of pages? In my tests there was always only one page of output that looked like a summary.
The main challenge here is that the models generally only output 4k tokens, so this requires saving the outputs and then combining them with a tool rather than using the model to combine them
Hi Sam, thanks for the CrewAI detailed walkthrough. Have you considered a scenario where the agent(s) need human input either for approval or additional guidance? Just a thought, I have no idea how I would do it.
🎯 Key points for quick navigation: 🛠️ En el video se muestra cómo construir un agente personalizado usando CrewAI, comparando los métodos secuencial y jerárquico. 🔄 Se utiliza un proceso secuencial para realizar investigaciones y escribir artículos, con un enfoque en registrar cada paso a través de funciones de callback. 🔍 Se emplea DuckDuckGo para las búsquedas, y se utiliza el modelo GPT-4 Turbo para optimizar costos y rendimiento. 📝 Se genera contenido en formato markdown, incluyendo un archivo de registro que confirma la finalización del artículo. 🧩 Los agentes involucrados en el proceso incluyen un especialista en investigación y un redactor, combinando herramientas de búsqueda y de consulta humana. 📈 Se establece un número máximo de iteraciones para evitar bucles infinitos en el proceso. 🔄 En el método jerárquico, se implementan agentes adicionales para obtener el tema de investigación de un humano, pero presenta problemas para integrar la entrada de forma fluida. 🧠 El proceso jerárquico permite la repetición y clarificación de tareas, pero puede necesitar ajustes para mejorar la comunicación entre agentes. Made with HARPA AI
I'm relatively new to this, but even as accessible starting off and learning is, I have some very basic questions I can't seem to figure out like...how do you manually type a string into an already running function to supply human input data for the agent? Are you actually running a function similar to: # topic = ["topic"]... #...{topic} Thank you and sorry for the essay lol As accessible as it is to start, you're going to get a lot of people that don't know ANYTHING asking basic stuff like this... So sorry if it's an obvious question and thanks again.
@@samwitteveenai yes sir, it was in this one. I'm not understanding how giving the human_tool to an agent allows me to type text so that the agent receives my inputs. Are you replying to the agent live? Is it a just a regular function that you run before you run the actual crew? Are you typing in that little area where the agents are doing their thinking? You never show how you actually did it...and I'm too new to have any kind of intuition or knowledge (knowledge you would consider common) , to figure out how the tool functions.
@@samwitteveenai With someone like your background, you definitely have a bigger vision of what you can do with such framework and how can it be implemented to our daily lives. I've always said the same thing back when you used to do a lot of LangChain videos. You could perhaps add more spices into it where we save everything in a database and maybe we can query that database etc.. Implement something like an agent doing RAG that has vision capabilities and gives output to another agent to do some other task
so grateful for these carefully crafted walk thru's, the accompanying notebook, the detailed but concise narratives, simply fantastic Sam !
🎯 Key Takeaways for quick navigation:
00:00 *🤖 Introduzione ai metodi di creazione di un agente personalizzato*
- Panoramica sui metodi sequenziale e gerarchico per sviluppare agenti personalizzati,
- Vantaggi e svantaggi di ciascun metodo.
02:28 *🛠️ Strumenti e modelli per lo sviluppo di agenti*
- Utilizzo del modello GPT-4 turbo e strumenti di ricerca gratuiti,
- Implementazione di funzioni per salvare i contenuti e comunicare lo stato al termine delle operazioni.
04:28 *📝 Processo di ricerca e scrittura con gli agenti*
- Descrizione dettagliata degli agenti di ricerca e di scrittura utilizzati nel progetto,
- Interazione tra gli agenti e l'utente per definire gli argomenti di ricerca e produzione di contenuti.
06:00 *🔄 Integrazione e passaggio di informazioni tra agenti*
- Uso della memoria condivisa e del callback per tracciare le operazioni degli agenti,
- Definizione di limiti alle iterazioni per prevenire loop infiniti.
08:00 *🎯 Definizione dei compiti e aspettative dei risultati*
- Assegnazione chiara dei compiti agli agenti e delle aspettative sui risultati,
- Importanza della specificità nell'impostazione dei compiti per evitare ricerche indesiderate.
10:02 *🗣️ Interazione iniziale con l'utente e inizio del processo*
- Avvio del processo con la richiesta di un argomento specifico all'utente,
- Primi passi nella ricerca e nell'elaborazione dell'argomento richiesto.
12:54 *📊 Revisione e ottimizzazione del processo basato sugli agenti*
- Analisi dei passaggi effettuati e del feedback ricevuto per migliorare il processo,
- Considerazioni su possibili miglioramenti e modifiche al flusso di lavoro degli agenti.
15:18 *🔄 Cambio al modello gerarchico e problemi correlati*
- Introduzione del modello gerarchico e dei problemi emergenti nel chiedere temi all'utente,
- Adattamento del processo per incorporare più agenti e affrontare questioni di ambiguità.
18:22 *🛠️ Risoluzione dei problemi e iterazioni finali*
- Risoluzione di problemi nel salvataggio e nella comunicazione dei risultati,
- Riesame finale del processo e delle prestazioni degli agenti.
20:50 *🚀 Conclusioni e considerazioni future sul progetto*
- Valutazione finale del progetto, considerazioni sulle prestazioni e miglioramenti futuri,
- Suggerimenti per migliorare la qualità e l'efficacia dei modelli utilizzati.
Made with HARPA AI
Hi Sam, you might also want to have a look at the context attribute in the tasks, it helps in keeping the crew on-topic.
Yeah that is a great point I totally forgot about putting that in here. Will include in a future vid.
Hi thanks for the walkthrough. Would you share the total cost of these crews? Also, I have a feeling that this research project may not fully reveal the collaboration capability of the agents. The final output is seemingly a brief summary of the Jamba paper. Did you get to do a comparison with the summarization result using zero-shot GPT-4?
cost for running them quite a few times was $2.3 i didn't work out the per run cost. My guess is around 1/10 of that or less. The difference here in regards to the paper is it did its own search and found the blog post and used it. I din't give it anything more than that. This was also a somewhat simplified example of what I would normally do too, to show the basics etc.
In regards to cost Haiku is way better for that.
Thanks, Sam. This is informative. That's much less than I expected. I'll definitely test it out.
Just about to drop a version with Haiku and it is even way less than that.
Thank you Sam :)
Have you tried the hermes models when it comes to function calling? They are fine tuned for that purpose.
Excellent video and explanation. It's now on the crewai discord channel.
Thanks!
Do you have the link?
Hey Sam, loving the vids. Have you used the feature where you pass task outputs into other tasks via the task context variable? Does this make much of a difference? This can be seen on the crew ai website.
Thanks for mentioning this. I just added this for the next video using CrewAI and Haiku.
very good content. thanks for sharing. btw how do you compare crewai with langgraph?
CrewAI is an abstraction above LangGraph. It is much easier to get started but also much more limited in regards to customization etc,
Great I am waiting for Claude 3 - I stopped using GPT4.
lol give it a day or 2 and I will upload it.
@@samwitteveenai Thanks
I'd love to see the Groq API being used, as it is hugely cheaper than both and insanely fast. Mixtral has 32k token window.
@bourdainedepiment3962 agree. would love to see that as well. faster means potentially more iterations.
hi, thanks for your video. I'm wondering if you know how to make crewai's output long and detailed? Maybe dozens of pages? In my tests there was always only one page of output that looked like a summary.
The main challenge here is that the models generally only output 4k tokens, so this requires saving the outputs and then combining them with a tool rather than using the model to combine them
Love your videos buddy. Have you used Agency Swarm? I'm curious to know if this is better or how it compares.
Thanks for the kind words. I haven't checked out Agency Swarm I will take a look at it and see how it goes. Thanks
Hi Sam, thanks for the CrewAI detailed walkthrough. Have you considered a scenario where the agent(s) need human input either for approval or additional guidance? Just a thought, I have no idea how I would do it.
Yeah totally Ken you can do that just like I showed but put it a a task later on in the process.
There is a custom tool for human input that LangChain which I have. Let me know if you want it.
@@matthewtschetter1953 thank you for the offer. Yes, please share it, I’ll take any help I can get.
Share your github handle or email
🎯 Key points for quick navigation:
🛠️ En el video se muestra cómo construir un agente personalizado usando CrewAI, comparando los métodos secuencial y jerárquico.
🔄 Se utiliza un proceso secuencial para realizar investigaciones y escribir artículos, con un enfoque en registrar cada paso a través de funciones de callback.
🔍 Se emplea DuckDuckGo para las búsquedas, y se utiliza el modelo GPT-4 Turbo para optimizar costos y rendimiento.
📝 Se genera contenido en formato markdown, incluyendo un archivo de registro que confirma la finalización del artículo.
🧩 Los agentes involucrados en el proceso incluyen un especialista en investigación y un redactor, combinando herramientas de búsqueda y de consulta humana.
📈 Se establece un número máximo de iteraciones para evitar bucles infinitos en el proceso.
🔄 En el método jerárquico, se implementan agentes adicionales para obtener el tema de investigación de un humano, pero presenta problemas para integrar la entrada de forma fluida.
🧠 El proceso jerárquico permite la repetición y clarificación de tareas, pero puede necesitar ajustes para mejorar la comunicación entre agentes.
Made with HARPA AI
Thank you!
can i use the Gemini API instead of the Openai GPT one?
I think the prompts will require a bit of a rewrite, don't think it will work straight out of the box.
thanks! :)
Can you also use gpt 3.5 for it? I try to figure it out but cant get it to use 3.5 model
Yes but 3.5 will often not work well for agentic behavior.
I'm relatively new to this, but even as accessible starting off and learning is, I have some very basic questions I can't seem to figure out like...how do you manually type a string into an already running function to supply human input data for the agent? Are you actually running a function similar to:
# topic = ["topic"]...
#...{topic}
Thank you and sorry for the essay lol As accessible as it is to start, you're going to get a lot of people that don't know ANYTHING asking basic stuff like this... So sorry if it's an obvious question and thanks again.
thats a good question. You can do it via the Human tools, which I think I used in this video or the earlier one.
@@samwitteveenai yes sir, it was in this one. I'm not understanding how giving the human_tool to an agent allows me to type text so that the agent receives my inputs. Are you replying to the agent live? Is it a just a regular function that you run before you run the actual crew? Are you typing in that little area where the agents are doing their thinking?
You never show how you actually did it...and I'm too new to have any kind of intuition or knowledge (knowledge you would consider common) , to figure out how the tool functions.
can u please continue this
what more would you like to see?
@@samwitteveenai With someone like your background, you definitely have a bigger vision of what you can do with such framework and how can it be implemented to our daily lives. I've always said the same thing back when you used to do a lot of LangChain videos. You could perhaps add more spices into it where we save everything in a database and maybe we can query that database etc.. Implement something like an agent doing RAG that has vision capabilities and gives output to another agent to do some other task