AI Agent 路线图: 独家对话 Make AI 负责人「Make.com's AI Agent Roadmap: Interview with Head of AI」 | 回到Axton

hi everyone welcome back to my Channel today is very special because we have SE with us he is a head of Applied AI at

make.com for those who are knew I am Exon a expert in AI optimization and automation workflows I teach make in my

air automation course hundreds of students have learned to build AI workflows with make.com today we will learn about

makes AI strategy s thank you for being here could you tell us about yourself and make.com yes sure hey thanks a for

hosting me here today um really happy to be here and also share a bit with your community on what we plan to make and also get your feedback onto it so yeah

I'm s on head of blood AI I'm leading our product efforts and our internal AI transformation I make and I'm a with my AI team and other leaders trying to

build the best a experience bring llms into make as well other tools around using LMS within automations so yeah let's let's dive into your questions so

uh let's talk about air features I see many interesting things happening on the air and automation area uh firstly zapia

is doing these things they have ai actions that work with more than 6,000 apps and 30,000 actions they made an AI

chatbot for customer service they have Z Central where teams can create a agents I think the most powerful air approach

are the chatboard and the zp Central and the n8n has this air features they have six different types of a agents they are

a agent has building memory and two calling capabilities and there also many other AI agent platforms such as

co.com as an AI agent building platform helps people make chat boards and Ai workflows and the most important is they

can publish them on many other platforms and the def. is a similar platform which makes real a or ition R engine for AI

agent and workfl I see all these companies focusing on AI agents so what is mies

plan to compete with these competitors we we are already competing uh with these competitors because we see uh that a lot of customers actually

building uh simple and reflex agents uh already today on make so there's like these five types uh of AIA agents so um

when you think about it uh what you what you can do uh uh already or uh today so yes it is hard to build goal-based uity

based and learning agents um but learning agents you could do with data stores so when you're saying um hey NN has their memory we have a data store

you know uh saier doesn't have a data store NN doesn't have a data store per se there you can do some uh database logic I I know but uh you you can do

this inside make already you can bring rack uh to your automations we have hundreds of customers or even thousands of customers actually using pine cone

and in their workflows so what you can achieve in um n8n agents you can achieve and make it is today probably not as

straightforward when you have like tool calling or things like that that's for that's for sure we we are aware of this um but when we see also like a lot of

like the LinkedIn posts around what customers are building we always ask ourselves do they need this do they need to have the tool ping capabilities or do

they just use a module before and have basically uh a predefined workflow and I had last week a long discussion around this where the customer said well actually we don't want to have this kind

of Freedom really like that and make we can predefine um what the llm and what the agent then behind the scenes uh should do so in in this regard we are

observing the market and we are aiming to build the right Integrations and extending our platform network

connection is unstable okay s is offline are you still here okay so basically we uh see that they're building with pine con and a lot of AI agents already today

so our goal is to bring the platform capabilities to work as seamless as possible with AI as fast as possible because we we don't believe that AI will replace people it will make your

building experience faster this is one piece on the other side we of course exploring agent capabilities to extend our platform capabilities to become

agent first capabilities um and with this motion we can tackle nearly all of the big right now emerging agent builders if you dive deeper into these

agent Builders they are not really workflow engine because they're legging a lot of things would be established over the years and when thinking about

sap here um sep here tries to go into this front end motion and database and table motion we're serving it for for quite some time now um ultimately it is

a strategic decision to say hey we want to expand into like a horizontal motion and cover more and more um tool capabilities this is something uh what

we are observing and we find it really interesting to see how how zapier extends there I think zapier may be

trying to find a right approach to build AI maybe it still on the way to finding what is the right things they should do

so how do you think about the n8n I personally I like n8n a agent right let's let's take the let's take the uh

agent of um n8n so in n8n um if I recall it correctly you have the possibility to use tools um and well these tools are

just modules or make you can have them previously in your scenario you we see a lot of customers actually having an AI

module where you say hey this is to goal return based on this goal as Jason structure and then they use l routers

and filter if the next step should be the execution of for example get data from LinkedIn get data from YouTube get data from Tik Tok and then they merge it all back together in another LM call and

they kind of like you you can build this and make it is not as represented in there in our front end because it's our

make unique visual logic which a lot of customers really like but I I get the the the Viewpoint of yeah I would just like to do everything in one module like

and it ended with their front end representation what they I think really nicely did is having these Atomic executions um it's basically like

allowing the agent to use one API call this is uh I think really interesting and um it's it's going to be interesting to see how they continue because

ultimately they're agent is a l chain wrapper you can just see it in their GitHub source code bases uh ultimately they just wrapped lank chain um initially it was even called Lang chain

um so this is really really interesting to see um how it evolves um I mean they have the custom code step this is something we are offering only in

certain pricing tiers which they offer like right out of the box so this is also something where we can evaluate that this is a approach that's feasible

nevertheless we are thinking um about how can we make it easier to have goal based agents because this is something where we see room to make it easier you

can build it today on make but you might need to make it easier yeah I think that the most difference between make and N

Mak AI module is are more like a w of API it's so even even make a integration

integrated more AI modules like perity chity and Cloud Etc an is not just a API

rep provide you a platform to to build an agent for example if I want AI to Bas my question and to search my notion

database or maybe search from the from the website from the internet the a the a agent can do this and makes a decision

between the between one AI agent modules because it has a memory and it has a two coding capabilities but in make I have

to build all these steps by myself and I have to deal with different API different apps as make oing gives me a a

easy way to use API make doesn't combine some things like a whole AI agent should do yes um yeah again we we try to make

it easier to build these AI agent capabilities um as you said you can build all of this on make it is not so easy you can build memory you can build

in spreat IDs you can build theoretically retrieving data from rag um yes the thing is it's not as straightforward as easy as it is in

these other platforms we we see this we understand this nevertheless we see a lot of customers that actually like our ease of use of just providing context to

LMS yeah um so we see that they actually build their rack flows with pine cone do this because uh they like the chaining of things so and if we compare it like

to Diffy I actually playing around Diffy quite some time so they struggle in getting proper integ ations and building the setup around the

Integrations so this is something you're never going to make it right for like all the customers that's that's for sure but we got to make it definitely easier

when it comes to the agent goal-based approaches goal based approach can you explain more yes so what you're what

you're basically describing in uh the AI agent capabilities is that you say to the agent do XY set let's take a super

simple example let's take the example of make me a pizza the agent has now five

tools it has recipe stove delvery heat and ingredients let's say so and it can choose what it needs to make the pizza

it probably doesn't need the delivery because you want to cook it yourself you don't want to order a pizza so you can skip the delivery so the agent chooses out of these five tools just four

tools and makes yourself a pizza let's say and this is a goal based approach you define I want to eat a pizza I want to eat a self-made pizza so please make

me a self-made pizza and the agent will do can you build this today and make yes is it easy not as easy as we would like to have it I I hope this explain it yeah

I understand the go based approach as you said but I still cannot figure out how can I implement this in make

currently how you can uh implement this in make currently um with the with the gold braed approach you mean yeah or

with h okay sure sure so when it comes to goal based approach you define yourself a goal in in llm you want to use let's say open Ai and for example

you can use open AI assistance and you can build with book and responses open assistance there's some YouTube videos around this there's some explanation

around this um or what you can also do is let's take an trophic where you don't have the assistance API in an Tropic you can again Define your goal and you can

say hey return me ad Json if delivery is true you basically Define the tools in ad Json and then you parse the Json and you route based on the Json is this nice

probably not you can acheve aeve it but it's not a greatest St yeah that's my question you make I can call an model

like CHT and tell the Char I need a pizza Char maybe give me different answers ask you to tud the answers by

myself I need to add a rotor like if chb said yes you can get a pizza I will go to and work Flor if CH said no I will go

to the other FLW but what what an AI agent can do it can do this All by

theirs yes and and that's the thing you don't see the reasoning you don't see the reasoning why it did XY set behind

the scenes and what we get from customers is that they like to define or manage the reasoning steps in between

mhm and and this is something where we might change we we are fully aware we

might change that we allow the model to reason and to execute till it's done with the tool

capabilities but you can also use your predefined routed tools let's say like hardcoded tools this is something we're fully aware and where we are open to

change okay that's great it also talk about the publish currently if I want to build a chatboard in make.com I have to

build a scenario and call this scenario by maybe web Hook from my front and application make doesn't provide a easy

way to publish the app publish the scenario right so sorry um so you're basically saying

um you mean like a like an API like scenario to API or yes scenario to API or scenario to uh web SD keyr something

like that okay so what you can do is you can do V hooks yes we are working on

authentication authenticated VB books um when it comes to VB books um it's not an API per se um

but connecting to front end and integrating it you can what we see customers building custom gbts with FB po we see that customers are using a lot

of FB Poes to integrate into front ends Yes again it's it's not as easy as we wish to have it um but there's so much infrastructure

and so um many things what we need to take in account if we change this if we change the approach there where of course the uh other platforms might be leaner might be leaner from

their customer basis to adopt faster but yes we all aware and we want to make it easier in the future to have uh scenarios as apis okay let's talk about

that feature plans uh what new features AI features we make haveon and can you share something yes I can um so when it

comes to AI we are exploring the world of um agents yeah and uh as a lot of customers saw on make waves we are

exploring um a simple way how you can interact with scenarios and also a way how you can turn scenarios into tools for llms these are things we are

exploring um as like as I'm as I'm leading product efforts I I'm I'm not going to can uh commit to timeline nor to exactly what we can do and will do

always AIT of like iterative uh development uh process so these are the things uh you can expect um you can also expect pretty soon uh the lounge of human and the loop especially made for

AI purposes this is something we are fully aware and want to push out because we see that this is something where a lot of customers uh can actually leverage content creation to a Next

Level okay thank you currently make has over uh 1,600 Integrations right is this

number correct or maybe more this number is sadly outdated at over 2,100 so oh

2,000 right okay great let me let me maybe ask one one question what would be uh for your community what would be the things you would like to see that make

does what are the things that come to your mind what we should do what would make your life easier um I'm super happy to explore this so uh if they can post

this in the comments or if they can share this uh for our make make feedback this would be really really generous and great uh because we want to develop the

capabilities that make our users successful okay I said co.com or def. do

you do you know these two platforms yes that's great the uh I personally like the two platforms appro for the AI features

which makes we we building air applications very easier for example they have building knowledge base they have building memories

databases and even they have code modules M does not have has a building code module if I want to execute code I

have to use third party apps yeah yes yes we we are aware I mean are we local are we no code are be Pro code that's

always a question um we are evaluating our users so yes again we are we are open for Change and we are open to adopt just changing market conditions and

changing interests in customer bases um yes so aent thanks for your for your feedback there and yeah keep it keep it coming because this gives us a chance to

become a really truly first AI agent platform or enabling it agents to be built on yeah so noode I think noode is

is our future but currently even even zapier has more than six 6,000 apps Integrations they cannot do

anything without a code so there's still something I we have to to use code but this is not very important the important

since are easy to build an AI boat and also easy to publish I think this is a two very important things yeah again

like we we are seeing uh these tools coming up and we are evaluating our capabilities where we can can go to um and yeah ex thanks thanks for your

interview okay thank yourself for sharing all this information you hey thanks everybody see you byebye