0:12

Hello everyone and welcome back to the Pac-Man stage.

0:19

Thank you for still humoring me with the clapping after I say that because I know if you've been in this room for the rest

0:26

of this morning, that's what I've been doing every single time before each speaker. So, a warm welcome back to this

0:32

stage. How does everyone feel after their coffee break? Do you are you feeling a bit more energized?

0:38

Yeah. Great. Good to hear. I know I do. I don't actually like coffee, but I got myself a cup of tea, so I've got my

0:44

caffeine fixed now. ready to go for the rest of the morning. But I am going to hand you over to our next speaker who is

0:52

Dr. Jason Hillary and he is the co-founder and chief technology officer at Zerve who is one of our sponsors for

0:59

today and that Zerve is a next generational development platform purpose-built for code first data

1:06

science and AI teams. So I'm going to hand it over to Jason. All right. Thanks very much. Yeah. So um yeah, so I think

1:12

that was a very good introduction. So I'll skip introducing myself. Um so very excited um to be here today to talk

1:19

about a topic that's um very fastm moving and also hopefully it'll be um uh

1:24

very interesting and if it keeps moving at the pace that it is uh hopefully most of the opinions that I have today will

1:30

change in a month and it'll be like advances in the technologies the tools that we're talking about and everything like that. So it's a a very exciting

1:37

kind of a time for I will say being a developer in in this space because of all of the the new tools and the the

1:44

velocity that it's moving at. So today's talk h is specifically around human agent collaboration uh and kind of

1:51

focused specifically on data and AI. So there's been a lot of um we'll say tools that are developed uh for developers

1:58

with the advent of like co-pilot cursor and everything like that but there's slight nuances and changes um required

2:05

if you want to kind of achieve those kind of levels of like automations for data science and and AI development. So

2:12

I just want to give a quick shout out. I did have a collaborator uh when creating this um presentation

2:19

and it was deserve agent. So um we released an agent during the week and um uh it was a busy week so we launched it

2:26

on Tuesday. It was short in time getting the slides ready. So I was like okay the talk is on human agent collaboration.

2:32

Can we work together to to whip some slides together? So whenever you see an asterisk beside uh some of the content

2:39

in the in the slides, it was fully generated in a in um we'll say a

2:44

PowerPoint presentation generated by uh an agent that wrote Python code to

2:49

generate a PowerPoint. Um so and we'll get into it and we'll kind of show show

2:55

that um we'll say uh in the in the live part of the demo and that'll be interactive. So, um yeah, I'll ask for

3:02

people to shout out prompts and things like that um kind of when we when we get to it as well. Um so, we have um one of

3:08

the the team from Zerve in the back who will surely shout up if nobody else does anyways. So, um uh yeah. So, kind of

3:13

we'll get um we'll get going. So, uh just 30 seconds I guess on Zerve uh to

3:19

kind of like pre preface it in terms of um we'll say when we get into the demo. So, what it is and what we've built is

3:25

the operating system for developing and delivering data and AI products. Um, so it's kind of to shorten the gap for I

3:33

will say getting things into production. So it's uh first of all it was kind of built specifically for human human

3:39

collaboration. Um so it was the only kind of collaboration a couple of years ago at the time but now there's um this

3:44

talk is about human agent collaboration and there's um it turns out some of the things uh that make humans work together

3:50

is also things that make humans and agents work together. Uh well um h it

3:55

was purpose built we'll say to be code code code based so there was a lot of tools and uh we'll say in the last uh

4:01

decade that were like no code low code kind of a tools for like data and AI um

4:07

we'll say and we didn't believe that that's where the the true value was um and now with the advent of agents they

4:13

can help to write code to create more customized kind of solutions uh with you we'll say and then um we also have a

4:21

number of customers that have been building like AI solutions with the the likes of OpenAI, Bedrock and Tropic

4:27

Claude. Um so some of today's topics are informed by that and other parts are informed by uh our own kind of

4:34

development work and trying to create a data NAI development agent we'll say. Um

4:39

so I'll keep the the background very brief and we won't start um at a typical point where we go back a couple years

4:46

talk about chatbt. We'll kind of fast forward to uh being more um we'll say just code ccentric and developer centric

4:53

and uh even more specifically on agents that have kind of like um kind of um

4:58

we'll say uh risen up uh over the last we'll say year or so to like um create

5:04

some of the like uh most popular kind of fast growing companies that the the like the world has ever seen effectively. Um

5:10

so I've just taken two as an example. Um so there's many many others um but it's

5:15

just to kind of give a a kind of an idea of kind of the different directions that these kind of a agent kind of coding

5:21

applications kind of um have found uh a lot of uh we'll say success. So cursor

5:27

is one of them. It's an IDE that effectively is like um an alternative for like a VS code. So traditional kind

5:34

of a code editor with like uh a AI kind of built into it from the from the

5:39

ground up and it has composer which is a function that can kind of do uh full end toend automations of uh software

5:47

development. It can run the terminal commands it can install uh npm packages uh everything like that. Uh on the far

5:54

side of the spectrum then is like lovable and lovable is a a tool if people haven't used it already that um

6:00

effectively uh has the promise of and does deliver to be fair on being able to

6:06

take a natural language prompt and generate a react application front end. So no coding uh knowledge required and

6:13

you can get a a full kind of React application. You can store it to git and you can take it and kind of make

6:18

modifications but everything you do is through natural language. So um both of

6:23

these have been like hugely hugely popular amongst like other tools as well. Um and kind of uh just to to kind

6:30

of highlight that they're um their popularity and their like people's willingness to pay for these kind of

6:36

tools is uh we'll say a good indication that there's a lot of value generated by agents doing software development. So um

6:44

people don't do it and kind of use it at scale for for for no reason or at least I I don't think they do. there's a bit

6:50

of hype around it but um I think they do generate uh like a significant value. So

6:56

when you talk about data NI kind of uh specifically uh there's kind of like different kind of challenges um that you

7:03

get presented with compared to if you were doing like traditional software development. So it's kind of like there's a different mindset you're

7:09

working with uncertainty to a higher degree. Um so there's like a data source that you've got to we'll say work with

7:16

whereas when you're doing software development it's a lot more like transactional uh you kind of like know

7:21

what the inputs and outputs are you find them through APIs um but uh that's not the case when you

7:27

do like work with data so data can change schemas can change um so there's a lot more uncertainty and what that

7:33

means then for like agents that you're developing is that they they require different context so uh those tools that

7:40

we um had talked talked about before, they work strictly off of the the code base and the natural language prompts

7:45

that the the person provides. Uh when you're working with data, you have another layer to it. You have to know

7:51

the types of the data, the column names, everything like that, we'll say. So there's like a a load of metadata that's

7:58

generated by the code you run that needs to be kind of um considered for you to kind of take the take the next steps. Um

8:05

so that is the effectively the inputs of your agent have to now consider

8:10

something that's far more variable. Um there's also different outputs. So the kind of code that you

8:16

generate is typically a little more interactive. So it's less functional code sometimes more like a uh we'll say

8:23

like ripple and things like that. So it's um a bit of a a different kind of an output we'll say for being able to

8:29

chain kind of tasks together. And then um there's also just a different process. So, uh, if you wanted to go

8:35

kind of end to end on a full kind of like data and AI solution and you ask an expert, they'd probably tell you, we'll

8:41

look at the data first and we'll see see what it see what it looks like. Um, they wouldn't be able to map out maybe

8:47

exactly what all of the 10 or 15 steps you have to do to be able to uh, we'll say get to the end of a process because

8:53

it's exploratory. There's iterations in it. Um, so there's like lots of learnings you've got to do. So um if an

8:59

agent was to kind of help you with these kind of processes, it should also take a similar approach. Um so that's kind of

9:06

um some of the the learnings we'll say um that we've kind of come across and this kind of influences three different

9:12

aspects then of we'll say how you set up your your agents. So they have different

9:17

like inputs which is your like uh the context that you provided and the system prompts. uh they produce different

9:23

outputs and then they also have to have like different kind of planning strategies and um things like that

9:29

compared to some of the some of the other tools. Uh that it does have potentially one advantage in the sense

9:35

that um the scope is a little smaller than traditional code bases. So you can fit more of the context of a whole

9:40

project inside the scope of a an LLM we'll say um like context window compared to we'll say a an enterprise uh

9:48

software database or software codebase. Um so one of the natural

9:54

questions you kind of ask yourself uh when you start to think about human and agent collaboration is like the comparative strengths and so the

10:01

benchmark is uh basically um a human expert being able to do a task by

10:07

themselves. So uh there's a there's an alternative kind of um where if you're using a tool like a lovable or something

10:14

like that that um the the benchmark is that they can't code at all. So anything they're able to produce that's useful is

10:21

um we'll say is a big step change for for the individual h for people that are

10:27

practitioners and data scientists kind of employed in companies it's a productivity gains so it's um how much

10:33

faster can you complete projects how much better can you produce it how much quicker can you find mistakes and so

10:40

those are the the kind of like focus areas I think for initial kind of like agents for for data and AI is kind of

10:48

uh enabling the practitioners to uh we'll say do better work. Um so humans

10:55

are are extremely good um and still are kind of like fundamental to to the process for like a number of reasons

11:01

that we'll kind of touch on in the in the next slide. Um so uh they kind of

11:06

provide all of the the context uh they learn differently than the the LLMs

11:11

effectively. Um, and then they they also can do like they have the world view, the domain expertise, everything like

11:18

that. The agents are super at um they're they're quite good at writing code. So

11:23

they can write it like pretty much like a senior developer level and they can all do it. So you can kind of scale all of these kind of like workers up on

11:29

demand. They're all able to write code. Um, and they just have to make sure that they're working on like useful tasks for

11:36

the most part. Um, they're tireless effectively, so they'll work 24/7. um

11:41

which is like something that's like completely uh kind of different. And um there's uh super at like certain things

11:48

like pattern recognitions, being able to look at like large documents, extract kind of useful information to be able to

11:53

kind of inform the next code that they'll they'll generate. Um so uh kind of long story short, uh

12:01

agents I don't think have gotten to the point where they can like fully automate um kind of like data and AI kind of uh

12:07

development and humans are definitely still required. Um and this is kind of down to to a few things. Uh one of them

12:14

was when you're collaborating with a uh like an agent of any sort. uh what

12:19

typically happens is you use it a few times, you kind of learn how it uh interacts and you'll change your

12:25

behavior to kind of like get the most out of the tool. So there's an adaptability from the human side that

12:31

isn't there in the agent side effectively. Um so that kind of is where the the real kind of like efficiencies

12:37

come from is that the the person learns to use the tools and that's still still the case. Um there's uh the ways the

12:45

learning is kind of um interesting as well. So if you think of like human consciousness how they kind of like um

12:50

how everybody h kind of like continuously learns you have everything kind of like in context all the time you

12:56

can develop your world view um an agent can't do that effectively so it's all depending on the context that you

13:03

provide which is actually one of the the tricky parts in terms of like between steps and calls of an LLM exactly what

13:10

information do you provide it with that it's able to effectively mimic that kind of a that kind of a process

13:18

And then there's also accountability. Um so uh if somebody here decided to kind of like work on the weekend on a on a

13:24

project, um do you know it's a effectively they're kind of doing it maybe on their own time. It's kind of

13:30

like a learning. It's a a kind of a a habit or some something like that. Um if an LLM kind of or an agent decides to do

13:37

it, there's a there's a bill attached to it. So if it works for 24 hours, uh you'd be um you'd be you'd be in

13:42

trouble, I think, in terms of um the the costs kind of associated with it. So there's an accountability part that um

13:48

humans also provide in um in it. And uh the people that are best posed to kind

13:54

of um get the most out of agents I think are the experts. Um so it's people it's

13:59

kind of um if you uh know the processes and kind of know what you want to get out of it and know what's possible uh

14:06

you'll do a better job than somebody who who doesn't effectively. So uh this is kind of one of the limitations I think

14:11

currently of of agents is that they're not proactive. they're more reactive at the minute. So you kind of provide good

14:18

prompts, they'll give you good answers. If you don't provide enough context, they might kind of like uh lead you to a

14:23

suboptimal solution. Um so uh experts are definitely still um the people

14:29

that'll get the most the most out of it. It's just how you work might change a little bit in terms of if you're if you

14:35

have multiple agents kind of testing things at once. Uh your kind of day-to-day workflow is is a little

14:40

different. uh the gap has shortened uh in terms of being able to uh make like data science

14:46

and AI more accessible through kind of like coding agents and kind of like co-pilot chat UPT um that you're you are

14:54

actually able to kind of like get some kind of like uh automations kind of like automatically developed. It mightn't be

15:01

production grade. It mightn't be the the hard problems inside a company. Um but uh for example, the screenshot on on the

15:08

screen here uh was my uh mother writing her first ever piece of uh Python code

15:13

effectively using the agent. So uh she logged in on on Tuesday after it was released, asked um could I get it to

15:21

kind of plan holidays for me. I was like, "Ask it. We'll see." And then it was like um it kind of went searched the

15:26

web, got all of the information, put it in a markdown block, and she was like, "Can it create like a travel plan for me?" And I was like, "Ask it." And it

15:32

did. So it created a couple of Gant charts and things like that. So it was um it was interesting to see there was like a little bit of vibe coding from a

15:39

a six six year old woman. So it was a it was a good um a good a good a good day

15:44

for the the product I think in terms of it. Um so um so how experts are and can

15:51

leverage agents and kind of what we're seeing people kind of use it for uh the most effectively is kind of doing more

15:58

experimentations. Um so one of the things uh is when you're kind of dealing with your daily life is like

16:03

prioritization. So there's only one thing you can code at a time or you might be in meetings or there could be

16:09

anything else happening. Um whereas if you have agents at hand uh you can

16:14

effectively uh give them initial tasks they can do the initial evaluations and

16:19

uh kind of like uh do preliminary kind of like tests and you can have multiple ones of these kind of working all at the

16:26

all at the same time. Um, so it's just like more experimentations, faster development cycles, uh, faster access to

16:33

to data and um, we'll say API documentations. That's another thing that's kind of like particularly useful

16:39

uh, that we're we're kind of seeing people people use it for. So it's finding open source data sets, uh,

16:45

finding how to use APIs, what APIs you're able to do. So all of these tools can integrate with the the web

16:51

effectively, so they can like automate that some of the the research part of it. Um so that's another kind of an

16:58

interesting aspect of it from a a data point of view. And then um it kind of ties in with the the more experiments

17:04

but the parallelism within a task. So if you define a task as being we'll say having like multiple uh kind of parallel

17:12

processes um we'll say you can you can have multiple coding agents working on it simultaneously. So there's no reason

17:19

that it's like single threaded in in that sense. So you can have three uh coding agents all working together on a

17:24

single project and uh they can all work independently to achieve a goal and then they're able to we'll say spawn a a

17:31

fourth kind of worker kick off a fourth worker that aggregates the results and and so on. So there's like efficiency

17:36

gains just in terms of um uh uh development time or wall time. Um so uh

17:44

to kind of uh before getting into we'll say the the interface itself and um

17:49

everything like that to do a bit of like an interactive kind of a a demo there's just a few bits on terms of for context

17:56

like what the typical kind of a workflow looks like at the minute it's typically

18:01

uh the person asks questions and kind of leads the LLM and kind of fills its

18:06

context effectively through through a chat. Um the agent will uh respond and

18:12

then effectively it'll kind of kick off a kick off a task and do some like implementation of the code. Um the human

18:19

will review it, ask for refinements. Uh then the agent will iterate and then

18:24

that kind of is the the current kind of a loop. So there's kind of a you can get agents to do to do work for you

18:30

effectively. Uh but you're still guiding it, reviewing it, evaluating it and kind of like leading the the process. Uh so

18:37

today it's very much like session based and transactional. So you kind of like tell it what to do and it does it and

18:44

then you review it and you might get it to to do more work. Um so I think I think this is a significant step from a

18:50

collaboration point of view because it kind of like um establishes a foundation. Um so uh humans do kind of

18:57

like learn to work with the capabilities and the limitations of of the agents. uh that kind of builds trust. Um uh but I

19:05

think this can move like fairly rapidly to something that's a bit more uh we'll say uh proactive or kind of like um kind

19:13

of closer to true collaboration. Um and that that can be done through a a couple of different ways um that I'll I'll

19:21

touch on in a in a second. Um but effectively it it could be in the next six months that they do act more like a

19:27

a teammate than a than a chatbot effectively. So they can be a bit more proactive. Uh they can kind of like

19:33

learn uh kind of like um processes better kind of like build up their own kind of context that they can reuse and

19:40

uh give real-time feedback. So there's a couple of tools already that do like Slack integration so they can like ask

19:45

questions you can interrupt uh chat. So without having to stop a chat

19:51

effectively to kind of like if it's going wrong h you can send a prompt in the middle it can readjust its plan and

19:56

continue. So it keeps everything in context and can kind of um we'll say uh be far more um as if you were working

20:04

with a with a with a colleague. Um uh the one thing I think they are missing

20:09

as well is the proactiveness uh which I I think I'll touch on in a second here.

20:15

So we'll jump into the the the kind of like the demo part of it in

20:20

a in a minute. Um so the uh but some of the fundamental aspects for like an interface uh for we'll say human agent

20:28

collaboration I I think is that it has to be visual still so it's like you can't like trust it to to just uh it has

20:35

to be like real time uh kind of and visual so you can see the work while it's uh progressing um so uh that's very

20:42

similar to how if you have two people working together Figma does it uh Mero does it all of those kind of tools so

20:48

that kind of an experience with with agents Um yeah, so that's the the real time

20:54

kind of aspect of it as well. Um there's like contextual kind of transparency. So it's like you're able to see the reasons

21:00

why it's doing an action, what it's what kind of a what it's about to perform is like important from a feedback point of

21:06

view. So it's not a black box. Um then there's like abilities to have like uh

21:12

interventions to be able to we'll say stop it, track it, and uh we'll say

21:17

adjust the plan on the on the fly. And then I think as well it's the ability to have multiple things working together is

21:23

where you actually get the um the real benefits of we'll say agents. It's like

21:29

being able to to multi-thread kind of tasks. Um is the something that uh as uh

21:36

people we can't particularly do. You kind of have to focus on one thing at a time. Um and you can have multiple one

21:42

of these workers will say all working uh simultaneously. Um, so I touched on some

21:48

of these before, so I'll just uh briefly run through them again. But the challenges with agents kind of is uh

21:53

very much that they're like fragile with their their context. So uh they can kind of do a very good job of planning out a

22:00

first set of steps. Uh but as soon as something kind of goes wrong, uh they can potentially get stuck because you

22:06

have to kind of like um give it enough information that will say uh it's able to have enough context in its next call

22:13

to like the LMS uh that it's able to uh find a path or a solution. Um so it kind

22:19

of has to build up context as as it's going and exactly what that context looks like isn't um isn't always like

22:25

immediately obvious. Um the communication then during the process is

22:30

a is a big one. So being able to like interrupt, talk, have it send you a message, ask for advice, uh things like

22:36

that is kind of um I think where h nearly has to go effectively to to kind of like um uh really be like um somebody

22:45

like a person that you're collaborating with. And then it's also very kind of reactive at the minute. So um it uh if

22:53

you ask the subject matter experts, so um the the great thing about all of the

22:58

foundational models is that they're kind of like almost like domain experts across every every possible kind of like

23:04

application to some to some extent. Um so if you were to ask a subject matter expert a question, they'd answer your

23:11

question and then they'd probably give you a a solution or advice. Um more often than not, the LLM will answer your

23:18

question but not give you like an optimal solution. So being a bit more proactive in terms of it advising you on

23:24

how you do your your tasks and stuff like that I think is an important part as well of being able to get to the the

23:30

best results when you're kind of working with with data and AI. It being able to tell you like if there's like um what

23:36

the the potential pitfalls are. So like training a model is kind of one aspect of it but you can train a a model that's

23:43

completely biased and can kind of produce harmful results. So uh it's not just always the end result but also

23:49

we'll say the process and the governance around it. So um that's kind of uh some

23:54

of the the current challenges. I think all of them uh there are solvable uh

24:00

over the over the short term to to some extent and hopefully like things will uh

24:05

keep evolving. Um so I'll kind of uh jump in now I guess to um to show uh

24:13

Zerve as a as an example. of we'll say uh human agent collaboration. So the

24:19

product is up on the the screen. So I'll kind of uh just give a quick overview of uh what it is. Um so basically it's a

24:26

development environment. Uh it's a code base. So you have a multiple blocks. Uh

24:32

you can drag the the blocks in we'll say and you kind of have each one of them is code editor and you've got a we'll say a

24:38

code full kind of code uh code view as well. Um and you can chain these together to build any kind of like an

24:44

application you want. It's cloud-based and it comes with uh we'll say um

24:50

compute settings per block. So it can run fully serverlessly. This is an AWS one in this case. It can run on any of

24:56

the the clouds. And the idea of having it in the cloud is that it shortens the path to production. So where you develop

25:02

is also kind of closer to where you deploy and you're able to do schedulings, APIs, all of the kind of

25:07

deployment options um quicker and you're able to kind of use the fact that um serverless uh has like parallelism

25:14

inbuilt. So if you have multiple people or multiple agents kind of working together, uh they're all able to spin up

25:20

and spin down resources on demand. So you can go from uh no agents to 10 agents without having to like worry

25:25

about any kind of infrastructure. Um, so that's kind of the the basics

25:31

kind of of it. And then there's the um the kind of bit that we're most excited

25:36

about at the minute. If I can uh I might have to go half screen to

25:42

talk to this because I have a uh timer on the uh bottom of the screen. But um

25:48

effectively uh what we have here is the agent and the agent is fully aware of everything inside of um inside of Zerve

25:56

H and your project. So if you have uh files, if you've got uh database connections, if you've got anything like

26:02

that, it's fully aware and it's kind of it's in it in its context. Um so you can ask it about your uh your projects. So

26:08

this was kind of my uh brainstorming for for the talk. So I put in the first uh

26:14

five blocks. So this one here which is a project brief and then these four blocks with my initial ideas and then I asked

26:20

it to uh we'll say over a course of number of prompts we'll say kind of expand on certain ideas and it started

26:26

to search the web kind of put in extra markdown blocks run uh some geni blocks

26:31

to generate some like potential scripts. Um, so this is the agent creating a geni

26:36

block that uh will say it brought the ginger 2 template a prompt to generate some text that we could use as a as a

26:43

talk track and stuff like that. So this is just the the agent will say helping to uh create the um we'll say the the

26:51

presentation and then I asked it could it create a PowerPoint and effectively

26:57

uh what it did was it gave it a couple of goes and uh I don't know which one of

27:02

these is the the most recent but um so I've never used PPTX Python before um so

27:09

it's was kind of new to me to see what it would do um but uh effectively if It opens here. I'll bring it down onto the

27:16

the other screen. Uh it kind of gave a a

27:21

draft of like what uh what we could include in a in a talk effectively. Um

27:26

so this was kind of like fully fully generated by we'll say uh uh the the

27:31

survey agent and I'll kind of get into um an example of it now. Um so uh we'll

27:37

kind of do some uh uh vibe coding uh live. Um, so, uh, we can say something

27:43

like, "Hi, um, okay, we're at the, uh, data science, uh,

27:55

festival. I want to demoer

28:01

uh, live any ideas." And uh effectively

28:07

what it'll do is it'll start to kind of create a create a chat. Um so it has

28:13

different kind of modes. So it's kind of like integrated with web search. It knows fully the the context of uh we'll

28:18

say um okay do we see any of these that are uh interesting? Uh okay. So uh let's just

28:26

take okay and go with uh number uh two

28:32

and uh find any

28:38

uh data sources you

28:45

require. So effectively what it's able to do is if it needs to it can like uh do web searches it can move on to like

28:52

planning agents and um things like that. So here it's going to move into uh the

28:58

code planner. It's now the code planner has started a a web search to look for like open source data sets or data sets

29:04

we can use and it um will uh once it's complete it'll present a plan and how it presents its plan is that it kind of

29:11

takes um some we'll say ideas from software engineering. Um so it's the

29:17

idea is that so this is like relatively experimental from from our point of view but it was the idea of having

29:23

milestones. So asking the agent to create milestones for its project. So uh if there's a level of uncertainty about

29:30

the project that what it would do is it would create milestones and iterate on it. So it would see if it could complete the first milestone before performing

29:36

the second one. Um and uh so I haven't really read this but uh it creates

29:42

milestone tickets. So here's three tickets that have been created for this uh particular milestone and each of

29:48

those h will uh spin up or be assigned a coding agent. So the planning agent

29:54

creates tickets. The tickets then are every one of the tickets is assigned a coding agent. Uh so there'll be some

30:00

level of like parallelism then when the the coding is is happening. So it's just like a a go for it. Um so effectively

30:08

then what it'll do is it'll submit the the different kind of like tasks to the the workers. Um and then uh you can kind

30:16

of watch it in in real time. You can stop it, you can change it um and um

30:22

things like that. So uh it takes a couple of seconds for it. So it kind of like each one can uh start to now create

30:29

u create blocks on the the canvas. And um we'll say uh we'll see it kind of

30:36

like uh working uh away. So it kind of creates blocks, runs blocks, and is able to kind of like de debug itself as well.

30:43

And you can be clicking on here. You can like add your own code. You can edit the edit the blocks and uh everything like

30:49

that while it's um while it's uh while it's running.

30:56

So this is just kind of a a bit of a kind of a a taste of like we'll say how

31:01

you can kind of like use the use the agents to we'll say uh do like data science and AI workflows. Uh it's also

31:08

like fully aware of everything like if you were to uh upload a file for example

31:14

uh so I know I have a CSV called songs uh

31:21

CSV if I did it up here

31:27

uh if I could spell uh so if I take this for example I can

31:33

throw it onto the canvas it's my only CSV V. So like uh I can go okay I want

31:40

to uh see what makes songs uh popular. You

31:47

don't need to like mention it's a CSV file you have or anything like that. It should know from the the context

31:52

effectively. Um uh I haven't looked at the data. So do some EDA and

32:02

uh create charts something like that. Uh so what it'll do is it'll present another plan

32:09

so it knows that it's like a coding request effectively and um you uh yeah

32:15

get presented a plan it'll do another kind of a set of uh coding uh in zerve you can kind of like uh so this is a

32:21

canvas um so this is kind of like one project effectively and you can have multiple layers layers of different

32:28

different types but uh each development layer is its own kind of isolated context for the um the the uh agent so

32:36

it'll will only work and edit blocks will say on the the current active uh the active layer when you submit the

32:42

submit the task. So it's like a go for it. Um we'll say

32:49

um so this is kind of like a a bit of we'll say um some of the uh yeah where

32:55

we think the the future can go with it in terms of it. It's kind of like a a lot of it is kind of around like a

33:01

productivity working with the agents more so than um it been like fully fully autonomous but it can definitely be

33:07

something that you can offload um offload some of the the workloads to and things like that. And um it can surprise

33:14

you sometimes. So you can learn learn a lot around packages, how things work, what alternatives are out there and

33:20

things like that by kind of like asking questions because it can um we'll say effectively like search the web. it has

33:27

like context of other projects that people have done before and things like that. So, it's um it can be a a good

33:32

kind of a a learning tool as well. So, we'll see that it's a now kicking off. It's um uh should know that it's the

33:39

song CSV. It kind of found it. It knew how to load it from the file system and it'll kind of like continue to kind of

33:46

uh code code we'll say. So this is um just a a quick example of we'll say um

33:53

uh uh yeah kind of like using an agent for data science and AI development and

33:58

how it's like a bit different than if you were doing a traditional uh software development. There's about a minute left

34:04

before I think there's some questions. So um just very quickly on the the future I think there's a couple of

34:10

things that are are particularly interesting in terms of how like agents can change deployment methods. um we'll

34:16

say so uh development is kind of what we looked at there in terms of it can kind of be a productivity gainer. It can help

34:22

to build and kind of like help you through the development process. It could help you to like access data and

34:28

documentation much quicker. Um but I think there's a a couple of things that are really interesting that could happen

34:33

in the space around like uh scheduling. So typically you would have like have a lot of things that would be like

34:38

schedule jobs. Um potentially what you could do is you could change some of those schedule jobs to be schedule like

34:44

agent tasks or agent messages. So there's a bit more like dynamism in it. It's able to like search the web for

34:49

like uh kind of input like variables we'll say based on like web searches before it does a run and things like

34:55

that which is a an interesting kind of a concept. I think application building is a is a thing that's um interesting as

35:03

well. So the the promise of streamlining a lot of tools was that you didn't have to be a front-end developer. that was

35:08

like their uh main uh we'll say um attraction. Um it's the exact same value

35:15

proposition of a as a lovable but lovable gives you like a react application that's a bit more customizable. Um so I think there'll be

35:21

more and more uh teams uh trying to use and integrate will say lovable with the

35:26

the internal applications that they they develop um because it has effectively the same level of if not less overhead

35:33

than uh developing uh itself. Uh there is a question still around whether the maintainability is is possible if you're

35:40

not a React developer and things like that. But that's something that we'll probably kind of see over the next couple of months. Um Git is the bridge.

35:47

So I think uh a lot of agents uh can kind of uh come to the market and they should all be able to work together. So

35:53

if there's one that you um uh like for front-end development or there's one that you like for backend development or

35:59

for this particular thing, h they should all be able to edit the the same code bases effectively. So you should have a

36:04

freedom of of movement between all of the the tools and git um and version control could be a good uh mechanism for

36:11

it and then uh the coding interface will evolve. So as you have more and more

36:16

agents kind of like working and delegating it's like um I I think there's I don't know exactly what it'll

36:22

look like but I think there'll be like more integrations with Slack more towards like natural language for some of those things. It asks you questions.

36:29

You're able to kind of like give responses while you're like doing something. you don't have to be at the laptop to kind of like unblock it and

36:35

things like that. Um so those are kind of some of the things that I think uh will will happen over the next uh little

36:41

while. And then finally um if anybody wants to to use the Zerve agent that was

36:47

uh we'll say kind of demo today. It's totally free so you can log in and just use it on on the app. I'd love to hear

36:53

we'll say feedback on it. So it's um we'll say yeah you can just sign up and and use it. And if you want a demo, uh

36:59

we'll say there's a booth downstairs as well that we can kind of go in. So if there's questions afterwards, uh we'll

37:05

also be downstairs to kind of like answer any questions or if people want to talk about agents data or anything

37:11

else. Yeah, happy to happy to to go through

37:21

it. Okay, so we actually have time for two questions. So, does anyone have any

37:26

questions for Jason? I saw a hand go up here first and then a hand at the back. So, I'm going

37:32

to do those two. So, I'll just come around with the mic just so it's on the recording.

37:38

Hi. Um, first of all, um, really good speech, a really good talk, and your

37:44

product seems amazing. Serve looks really good. And based on that, I I

37:49

couldn't help but think, do you think there's like a bit of a delusion in the data industry where we're convincing ourselves that we're still necessary?

37:56

Um, I I don't think so. Not yet. Anyways, it's in in the sense that it's um I do I do think it might change in

38:03

terms of like how how people work. Um I think like uh where where it can actually really really help is on the

38:09

production side as well in terms of um there's like lots of pipelines that fail silently and things like that that um uh

38:15

an agent could always be like watching monitoring and like look at looking at things. So um uh I I still think pe people are are

38:24

necessary at at this time anyways would be my um my feel. And if you yeah I'd

38:29

say you you probably if you like log on to Zerve try it out for a while you'll kind of see if you ask it open in the questions um you generally do have to

38:36

like narrow it and kind of point it in a in a direction we'll say. So um it's not quite at the point where it's um uh I I

38:43

think the knowledge is there in the LMS to be able to do it but the there isn't a system that I've seen yet that that

38:49

can uh will say um do it. And then I think there is also a part where humans

38:56

and how we learn is like fundamentally different like as the agent runs uh you

39:03

build up a like a world view of like the data set that's like uh probably fundamentally better than the agent that

39:09

was producing the code because it didn't learn anything necessarily. It kind of like used context and steps. Uh but uh

39:16

effectively if you were to ask to start again it doesn't have the previous experience. So it would make all the

39:21

same mistakes we'll say whereas the the person person wouldn't.

39:27

That's great. Thank you. And then we had one question down here. I believe it was you that had your hand up. There you go.

39:33

Uh thank you. Is Zerve better than Open AI's codecs? Um similar. Oh, it's um h

39:42

that's for people to try out and judge I think in terms of um h I'm biased. So I

39:47

I think we'll h if if you don't think it is yet, I think we'll we'll get there um in terms of it. But um uh I think

39:54

there's certain things that it does that is uh interesting um that's different from a organizational kind of point of

40:01

view or developer point of view in that it um uh it can manage the infrastructure inside the the company's

40:07

kind of account. So it's fully linked and self-hosted. Uh the free version is like SAS. You can just log in, try it

40:13

out and everything like that. But uh you can fully self-host it inside like a company's account. So it has access to to database fully secure. Zerve never

40:21

processes kind of data or anything like that in in that sense. Um and then there's also the way that the uh the

40:27

context awareness of the the data I think is another advantage of Zerve. So

40:33

when it runs a file it um will say h has full access to any of any and all of the

40:38

metadata about the contents that were also uh produced. will say so um code uh

40:44

agents typically uh focus on the code um we'll say but when you want to do a next step in a data project you do need to

40:51

know about the data we'll say so I think there are two two things where it's a

40:56

bit different at the minute that's great thank you so much Jason for

41:01

this amazing presentation as well it's been great and I was going to say if you have any

41:07

questions come and find Jason at the end and we'll be resetting the room for our next speaker in just a couple of minutes. So stay put as well. So yes,

41:15

see you soon.