Anthropic’s Blueprint for Building Lean, Powerful AI Agents

[@OfficialChatbotBuilder]

I just did a live yesterday showing how to build powerful agents with no-code 🔥

[@harisjaved1379]

This is 100% true! I build a multi agent app on azure and it was too complex, the goal was to take a user request and use tools to perform analysis, the problem was that the requests we not always sent to the right agent, best to use a pre-defined / predictable paths.

[@andrew-does-marketing]

Did you build an ai decision maker agent that acts as the initial layer to take in a request? It’s only training is how to make decisions on which tools to call.

[@harisjaved1379]

@ yes, there were 4 total agents, the first layer was the decision layer that passed the request to the next agent based on the initial classification and data

[@andrew-does-marketing]

@@harisjaved1379 what prompts were you using for the orchestrator. That’s where the magic happens. You can’t go budget with orchestrator. IMO you just need to use o1 and have a large pre-prompt with examples and directives in the chain of messages that are passed to the decision agent.

[@tollington9414]

I totally agree with this assessment- after all we are just calling a very simple API when working with LLMs - no need to learn a framework to do that

[@sun-ship]

Thank you so much! This was very clarifying.

[@JayTrainerBand]

🤯 Wow, this was so helpful and totally eye-opening from a non-developer perspective! I have a question for you: I've been watching so many KZbin videos where people say they are building agents, but if I understand what you're sharing here, they are really just building workflows and calling them agents. Is that correct?

[@vladyslavkorenyak872]

Yea these words are new and there is not a clear boundary around them yet.

[@miguel1795]

thats partly correct. think of an agent as a worker that performs a task based on a workflow

[@engineerprompt]

Yes and just to clarify it further. You can just use an LLM (with tools) in a workflow to do most of the things. You mostly dont' need an LLM to make autonomous decisions but rather use pre-defined workflows.

[@patruff]

Yes

[@patruff]

​@@engineerpromptexactly this, most people just want tools with their LLM. Agents are so buggy right now. Wait a year I'm sure they will be better.

[@FinanceGraphics]

Beautiful

[@aaagaming2023]

Its all in the prompts, and the use of prompt chains.

[@OfficialChatbotBuilder]

#PromptChainGang

[@barrettvelker198]

I have yet to find a prompt chaining framework that is not complete trash

[@orthodox_gentleman]

Same brother. I think this bigger companies may be intentionally trashing these frameworks…think Microsoft GitHub…I am realizing that everything seems to be a scam.

[@sumitsp01]

Which framework do you feel is best at the moment for building AI agentic solutions? I am learning langgraph which is descent, but I am hearing PydanticAI is great as well.

[@broformation6530]

Tbh I've come to the conclusion that raw dogging it is simply the best option

[@OfficialChatbotBuilder]

Sucks to hear. What has failed so far, like what’s the actual problem?

[@javi-to-msaas]

Would you try a NodeJs alternative?

[@sun-ship]

What do you think of pydanticAI?

[@mayuringole9679]

Hey can you tell me which application u r using for that GUI demonstration?

[@engineerprompt]

if you are talking about my screen recorder? I use screen studio.

[@verma.prashant]

I built this system 11 months ago; I am ahead of RAgentic 😂 time.

[@orthodox_gentleman]

LangGraph sucks unless you have a Mac I guess…

[@patruff]

I hate agents