Hey, how's it going? Maronei here from

CodeCloud and today we're diving into

OpenAI's agent builder. You know, the

so-called end killer that everyone's

been talking about. Now, for those of

you who haven't heard of agent builder

yet, it's OpenAI's new platform that

lets you build AI agents with minimal

setup. Think rag agents, custom tools,

and conversational AI all without

writing complex code. Sounds pretty

compelling, right? We're going to

compare agent builder and end across

three key areas: functionality, ease of

use, and maintenance. By the end of it,

you'll know exactly how Agent Builder

stacks up against Ned end and whether it

actually lifts up to the height. So,

let's take a builder for a spin. So, in

order to gain access to agent builder,

you want to go to

platform.opai.com/asian-builder.

So, this is different from the usual

chatg link that you want to go to. So,

once you get on there, depending if

you've already signed up in the past

before, if you haven't, what you want to

do is to just get signed up and you'll

be led to this dashboard where you can

start creating workflow. Now, what you

want to do is to just hit create. And

the first thing that you notice the

default setup within the environment

starts with two nodes which is the start

node which is a trigger node and the

Asian node. Now with the start nodes you

can see this is by default a chat input

node. So this is currently the only way

that you can trigger workflow. So if

you're used to nan's web hook triggers,

schedule triggers or email triggers.

Yeah, those aren't here yet. In fact, it

does seem like it's more geared towards

chatbot use cases because if you were to

go back to the dashboard here, it says

build a chat agent workflow with custom

logic and tools. So, you can sense that

this is purpose-built for one particular

use case which is chatpot, right? So,

it's not very versatile unless they

start developing new trigger nodes, new

way of triggering the workflow for

different types of automations. And I'm

sure OpenAI will be working on that down

the line, but for now it's pretty

limiting. So going back to the workflow,

what I want to point out is when you go

to the Asian node here. So you can see

it is pretty similar to what you see on

NN, you do have a assistant promp field

that you can type into. Currently it's

just a you are a helpful assistant. So

you can change this depending on your

use cases. But again, this is going to

be chatbot functionality. And within the

note, of course, you have the choice of

including chat history, which is pretty

similar to the memory context within

N&N. And you can choose the model, you

know, whether it's GPT5 or not. But I'm

going to leave it as GPT5 for now.

Reasoning effort is also something you

can specify whether you want it to be

minimal to high. And under tools, this

is where things get interesting, right?

So under tools, you can actually attach

a couple different tools. Obviously, the

biggest one here is the MCP server,

which gives you the capability to get

connected with external tools. However,

I want to point out that currently the

MCP functionalities are pretty limited

as well as buggy. And if you go into

just the Gmail MCP, you'll see that

there only two capabilities with the

Gmail MCP, which is search and read

emails. Now, of course, that's not very

useful because what I want an agent to

do is typically send emails on my behalf

and do other things, right? But one

thing to note is that it is relatively a

lot easier to connect the Gmail MCPS.

What you do is you click on get access

token and where you're led to is the

Google o page right here and you can

simply scroll down to Gmail API and over

here you can see that there are a couple

of functionalities or permissions that

you can grant. So in this case it has

the send option here but it doesn't work

even if you were to select it but I'm

just going to select that. When you

click authorize APIs, it's going to give

you the key which you can paste over in

the access token here. But I'm not going

to connect that because literally it's

not very useful if you could only search

and read emails at the moment. So it's

something for obviously OpenAI to work

towards. But again with the current

existing list of MCPS that are available

here, you can see there's only a handful

of tools and apps that are on here. Of

course, you have a choice of connecting

it to your MCP server. And again with

the limitations of Gmail, what you can

actually do is you can go to NN for

example and you can create your own MCP

server. This is a simple MCP server with

a Gmail tool connected which allows you

to send a message through Gmail. So it's

a simple MCP server. And what you would

do is you would just copy the URL here

and go back to the MSP server details

here and paste the URL in and fill up

all the other stuff currently on. I've

chosen to have no authentication because

this is just a demo. But then you can

actually toggle this to none. And once

you connected, you can actually get

access to send Gmail. However, the

obvious question is if I'm doing all of

this already on NAN, then why wouldn't I

just orchestrate the entire workflow

natively on Naden? Why would I even need

agent builder, right? So right now it's

pretty limiting in the things that is

available for us to use. But I just want

to show you how you can connect to your

custom FTP servers even ones built on

edit end. So right going back the other

tools that we see here that are very

interesting and the highlight of agent

builder at least to me is the file

search tool. This is where it actually

becomes a rack agent where you actually

upload context and information that you

wanted to take reference to or retrieve.

So for example here I'm going to upload

a standard operating procedure document.

So I'm going to upload the SOP PDF. And

how this PDF looks like it's simply a

document outlining the ticketing

customer policies of a fictitious

airlines called ANOVA. And it's just a

couple pages of document outlining you

know ticket refund reservation policies

etc. So here I want to name this

ticketing policies SOP. All right. I'm

going to name the vector store that. As

you can see there is an option where you

can select a vector store to attach if

you already have an existing vector

store. So that's really neat. And here

I'm going to just click attach. So there

you go. That's just how easy it is to

attach a memory context or a database to

the agent. Meanwhile, with Naden, what

you typically need to do is you first

need to set up your own vector database,

whether that's superbase or pineon

vector and probably you would have a

workflow that might be similar to this

where you upload your files on drive and

then that gets upserted into pineon

vector or you can manually upload or

upsert it directly from you know

superbase or pine vector store right and

that's step one and then step two you

might have something like this where you

have the agent attached to the pine cone

store like this and it's just not the

most easy to use mechanism as opposed to

just attaching it right on the tools on

the Asian natively on OpenAI agent

builder right like this. So to me that's

the biggest win that I see for the agent

builder. But again coming back to the

tools here there's also the web search

and code interpreter which both are

really useful web search. They can

actually go out to the internet to do

the search the way you would ask to go

out internet search a particular topic.

So in this case because I've already

attached the ANOVA ticketing policies I

just want to create a customer support

chatbot here. So to show you guys how

this would all work for ANOVA. So under

the system prompt, instead of saying

you're a helpful assistant, I would say

there are helpful customer support bot

for Airva, your job is to answer

customer queries

about ticketing by referring to attach

vector DB for the most accurate

information. So I'm going to give it

something generic like that. and we're

going to hit preview to try it out. And

for those of you who are just logging in

the first time onto OpenAI platform,

this may not be available. You need to

do an organization review. What it is is

basically just you uploading your ID and

then take a couple of photos of yourself

and that's pretty much it. It's pretty

straightforward. But preview is how you

would test out the workflow as opposed

to running a test run on any net. Right?

So here I'm going to ask what is my

ticket refund policy. Okay. So I'm going

to ask it something like that. So as you

can see it's searching the refund policy

details. It's actually using the file

search capability which is correct. It's

good dipping into the vector database

and getting me the most accurate answer.

So there you go. It's telling me the

refundability depends on your fair. So

it actually gives me the citation as

well from the document. Right now

there's only one document in the

database. So you know it's not really

helpful but if you have multiple

hundreds thousands of documents this is

going to be helpful to let you know

where it's getting the information from.

Okay. So one other thing that's pretty

cool is the guart rails node for options

where it can actually help you mitigate

some of the hallucinations and specify

the moderation and even filter out

personal sensitive information that

might go through the workflow. So this

is something pretty cool and one of the

wins as well in my boat. Under the logic

and data, it has the usual stuff which

is the if else while logic gate and the

user approval, you know, having human in

the loop which is great as well, but

already things that you can do on NN.

Well, in specific functionalities, it is

obviously not as mature yet as NN. NN

has been around for a while. It is

something that OpenAI can work on. But I

do want to touch on the structural

aspect which in the short term not going

to be able to be addressed. So one of

which is the fallback model. As you may

or may not know on endn you actually can

specify a fallback model where once you

toggle on the enable fallback model in

the AI agent you can actually select

another model for example the anthropic

chat model or other LLMs in case that

open AI has gone down your workflow is

not going to car out your workflow is

going to continue to flow with the

fallback model that you so choose. So

that's a huge structural gap because of

course using agent builder you are

relying on just one model which is open

AAI not just as a choice because

sometimes depending on your use cases

you might want to use anthropic and lama

instead of just open AAI but also as a

fallback right in case open AAI runs

into some issues you want to be able to

automatically toggle into another LLM so

that your workflow does not fail. So

that is a structural issue that is not

easily addressable. Another thing that

is not easily addressable is the fact

that edit is actually open source. So

you can self-host it or in any

environment of your choosing. And in

fact part of the course that we built

for Naden will guide you through on how

you can self-host Naden on top of a lot

of other things like foundational

knowledge and being able to build

workflows step by step. So if you're

interested in that, we're going to link

the course down below. So do check that

out, the NN zero to hero course. But for

whatever reason, if you were to want to

self-host for security or control, you

don't have that option here with Asian

Builder. Obviously, it's hosted by

OpenAI. So, you have to be reliant on

that. And of course, for the UI itself,

while it's super beginner friendly, it

doesn't show the input and output data

the way that NNN does. So, it makes

debugging and troubleshooting a lot more

challenging, not as straightforward as

compared to NN. So it does feel that

like N&N is geared more towards

technical and developer tool versus the

Asian builder is geared towards more

just specific chatbot use cases and also

super beginner friendly easy to use

segment. And on that note it also does

not have the debugging or error

notification capabilities that edit end

has at least not currently. So for

example in edit end you can actually set

it such that if the workflow errors out

it can trigger another workflow which

sends notification and stuff like that

so that the maintainer can be notified

and go into the workflow and be able to

fix things right. So right now that's

not readily available with agent

builder. So in terms of functionality

definitely end takes the cake when it's

compared to agent builder. Uh in terms

of ease of use, agent builder uh

definitely is much easier to use if

you're not used to NN and you're trying

to choose a P play no code tool. And of

course, if you're only trying to build a

chatbot currently, it is a lot easier as

well because it does have the in-built

chat kit. So you get a readyto-use chat

interface without having to build one

yourself. Pretty convenient if you just

want to deploy and share chat agent

quickly. In terms of maintenance, as I

pointed out earlier, it is pretty tough

to maintain if you don't know kind of

the granular details between the input

and output data of the particular nodes

as well as not being notified or having

another workflow triggered whenever an

error occurs on your workflows and not

having a fallback llm in case openai

goes down. So, you know, maintenance

might be an issue there. So, I'm sure

OpenAI is going to address a couple of

these things and continuously improve.

This is the first iteration. So, I'm

hoping to see much more improvements in

the next version. But yeah, so here's

the bottom line. Over the ice agent

builder looks awesome on the surface,

but it definitely has its limitations.

And I'm not nitpicking here. The

workflow trigger options make it useful

only for chatbot use cases, and it lacks

a seamless integration to external

tools, and even the MCP functionality is

pretty buggy and limited. So, it needs

some serious work. That said, OpenAI is

doing something pretty cool here. Give

it a few more iterations, and this could

actually become really powerful. But

right now, it's not quite the NN killer

people are claiming it to be. Let me

know what you think in the comments

below. And if you're interested to learn

more about NNN, I have a course for you

right here at Coke Lab, which not only

covers practical build, but let you

practice in actual real life and end

environment risk-free. So check it out

in the link below. And as always, if you

enjoy the content, hit that like button,

and I'll see you in the next one.