========================================================================
Get started on Sui through Agentic Engineering
Speaker: Raphael Chia — Mysten Labs
ETHGlobal Tokyo 2026 · Workshop · Friday, September 25, 2026
Video: https://www.youtube.com/watch?v=fkFQjrpOVUA
Transcribed with aloha by the AI ʻohana 🌺
========================================================================

[00:00] GM guys, it's about time, so I guess I'll start.
[00:14] Very good to be here today.
[00:16] My name is Rafael and I'm from Mr. Labs.
[00:19] I'm a solutions engineer and it's been
[00:21] about a year since I've been with Mr. Labs.
[00:25] My day-to-day job is basically
[00:27] either integrating solutions with partners,
[00:31] or I'm going to be right here talking about SUI,
[00:34] talking about interesting topics,
[00:36] interesting conversations with you guys,
[00:38] with builders like you guys.
[00:40] Today, the topic that I'm going to be sharing about
[00:43] is agentic dev with SUI,
[00:45] and more importantly,
[00:47] how do we get more quality and less slope?
[00:52] The reason why I'll talk about this is because Stanford
[00:55] did a landmark study of about 100,000 developers,
[00:59] and they found something very crucial.
[01:01] They have agreed that AI coding increases the productivity,
[01:06] despite all the AI-related errors and all the tech depths.
[01:11] A little fun fact, they also discovered that productivity
[01:13] in greenfield code bases is way higher
[01:17] than in brownfield code bases when using AI.
[01:20] I'm sure some of you guys can already relate to that.
[01:23] When you're building a lot,
[01:24] when you're participating in hackathons,
[01:26] a lot of the times you're dealing with greenfield code bases
[01:28] and obviously AI helps a lot.
[01:31] That's something we can agree.
[01:34] What is the goal of AI engineering, actually?
[01:37] Who actually thought of this?
[01:39] Who thought of what is the goal?
[01:40] What is the end goal?
[01:42] Now that everybody is using AI,
[01:44] personally, I feel that there's three goals,
[01:47] and the first is to solve complex problems.
[01:50] That is a very big goal, you know?
[01:53] And the second is to do a large amount of chores
[01:57] or work accurately,
[01:59] not just doing a lot of codes,
[02:01] not just outputting a lot of codes,
[02:02] but doing it accurately.
[02:03] And lastly, no slope, no tech depths, right?
[02:07] Because what is the point of, you know,
[02:09] creating, solving complex problems,
[02:11] doing a lot of output,
[02:13] and all you get is like,
[02:15] slope, it works in production,
[02:17] but are you confident enough to say that, you know,
[02:20] you are in control of your codes,
[02:22] you are in control of everything that's happening,
[02:24] you chose the architecture, you know,
[02:26] with slope and tech depth,
[02:27] you can't really say, right?
[02:28] You can't really say that, you know,
[02:29] you have control, you know your architecture.
[02:31] So that's, in my opinion, right?
[02:34] These are the goals of AI engineering,
[02:36] and there's actually one more very important goal,
[02:38] which is, I'm sure some of your,
[02:42] most of your Saudi strays before, right?
[02:44] AI telling you that you're absolutely right.
[02:47] Right, it's familiar, you guys.
[02:48] We don't want that, right?
[02:50] So this is one of the biggest goals.
[02:51] We don't want to see this,
[02:53] I mean, in fact, we want the AI to tell us
[02:55] when we are wrong, right?
[02:57] What's the point of, you know, agreeing with us?
[02:59] So these are the few goals of AI engineering,
[03:02] and how do we get closer to the goals, right?
[03:07] This is gonna be the main topic I'll be talking about today,
[03:10] and this is context engineering.
[03:12] And I'm sure, again,
[03:13] most of you I have heard about context engineering, right?
[03:16] It's a word that is floated around.
[03:18] It's quite popular.
[03:19] So the idea is that if the only way to affect
[03:26] an LMS output is through your context window
[03:30] and your inputs,
[03:32] why are we not spending more time
[03:34] to manage the context window, to manage the inputs, right?
[03:38] I'm sure we can all agree,
[03:39] because none of us can actually tune like Astra 6
[03:43] or Opus 5.5 on the flight, right?
[03:46] We can't do the parameters.
[03:48] What is the only thing that will affect our output?
[03:50] It is your inputs and your context window, right?
[03:53] For the normal users like us.
[03:55] So if that is the idea, you know,
[03:57] why are we not spending more time and effort
[03:59] managing the context window and controlling the inputs, right?
[04:03] So the entire thing is just based on this idea.
[04:06] I'm sure you guys have seen these words before,
[04:08] context-driven, context-road, garbage in, garbage out, right?
[04:12] And honestly, if we are the one generating,
[04:16] if we are the one prompting,
[04:17] if we are the ones using AI in our engineering, right?
[04:22] And all we do is we hope that the output is good.
[04:26] Basically, it's just like a casino, isn't it?
[04:29] We're just gambling.
[04:29] We're just hoping that we have a good output
[04:31] at the end of the day, right?
[04:32] But we don't want that, right?
[04:34] And speaking of about casino,
[04:36] I think, you know, prompting and casino
[04:39] has a lot in common,
[04:41] because in a casino, we buy chips.
[04:43] When you are prompting, you buy tokens.
[04:45] It's the same, man.
[04:46] You spin a slot machine, you press generate.
[04:49] It's basically the same.
[04:50] You might hit a jackpot.
[04:51] You might get a working code, right?
[04:53] In a casino, you just attract the lights and animations.
[04:56] Yeah, right, great idea.
[04:58] That's what you're getting, prompting.
[04:59] And casino always profits, cursor always profits,
[05:02] the AI provider always profits, right?
[05:05] That's so much overlapping,
[05:06] and this is what we don't want.
[05:07] We don't want prompting to turn into a casino.
[05:09] We want to be in control of our outputs, right?
[05:14] So very quickly, there's two ways
[05:17] that you can manage contacts
[05:18] that I'll share quickly about.
[05:20] The first is frequent intentional compaction,
[05:23] and the second is research plan implement, right?
[05:27] So we're gonna take a look at what are all this.
[05:30] So frequent intentional compaction,
[05:32] typically, the naive way to start prompting
[05:36] is to go through re-staring.
[05:38] So what is re-staring, right?
[05:40] Imagine you have a contacts window.
[05:42] You have a chat with your AI provider, right?
[05:47] You have a contacts window,
[05:48] and the output you want is not there yet,
[05:51] and then you will stand to see this,
[05:53] like, oh, no, do X mistake,
[05:55] or this is not good enough.
[05:57] Maybe do Y approach instead,
[05:58] or what is the difference?
[06:00] Use the best method, you know?
[06:01] As you can see, when you go through these kind of prompts,
[06:04] you're kind of re-staring the current contacts, right?
[06:07] Re-staring to make it right,
[06:09] and this is the simplest way to start prompting.
[06:11] I'm sure all of us, like,
[06:12] when we are doing debugging, right,
[06:15] all these are very familiar to us,
[06:17] and this is called re-staring,
[06:18] which is, it's simple,
[06:21] but what it does is it blows up your contacts window.
[06:25] So imagine you do this, like, five to 10 times.
[06:27] Now, in your contacts window,
[06:29] you have, like, five to 10 different wrong methods,
[06:32] and every time you send your next bubble,
[06:34] you send your next prompt,
[06:35] the entire contacts window goes together.
[06:38] So you can see how that starts to, you know,
[06:40] clutter your contacts window with inaccurate information,
[06:44] right, which is why I want to talk about
[06:46] intentional compaction.
[06:48] So what is intentional compaction?
[06:50] It's just as simple as prompting, saying,
[06:52] you know, write everything down to a markdown file,
[06:55] ensure to note down the approach,
[06:57] what do we do, and how are we gonna move forward,
[06:59] and know what is the current problem that we are working on.
[07:03] This is as simple as it is.
[07:04] There is no hard rule, there is no, like, fixed prompt.
[07:07] Basically, compaction is saying,
[07:09] throw everything we've done into a markdown file,
[07:11] I'm gonna move on to a fresh context.
[07:13] This is what intentional compaction is, right?
[07:16] And again, maybe you might ask, you know,
[07:18] frequent, how frequent?
[07:20] Basically, again, when you see this,
[07:23] it's probably time to move on.
[07:26] But jokes aside, you will know when to compact,
[07:29] when you start to code a lot,
[07:31] when you start to deal with, you know,
[07:33] a lot of AI-enabled engineering, right?
[07:37] Maybe three or four wrong prompts,
[07:39] you know, it's time to move on.
[07:41] You know, it comes with experience,
[07:42] it comes with context of the code base, right?
[07:44] So why do we need to compact?
[07:47] Because we know that right now,
[07:48] contact size is limited.
[07:50] What is the common contact size
[07:53] that we are getting right now,
[07:54] as a normal user?
[07:55] 256,000, one million,
[07:57] some of them have one million, right?
[07:59] It's standing with two million.
[08:02] I don't know, but do we even need that much, right?
[08:04] So typically 256K or like one million
[08:07] are the typical numbers and it's limited, right?
[08:10] And everything takes up your context window.
[08:12] Your file input, output, modification,
[08:14] everything takes up, right?
[08:16] Even two calls, especially two calls.
[08:18] If the two calls are not optimized,
[08:19] it takes up a lot of context window.
[08:22] So the goal is to keep your context window
[08:25] between 40% to 60%.
[08:27] And this is stated by Joffrey Huntley.
[08:29] Who knows who is Joffrey Huntley here?
[08:31] Joffrey Huntley, some of you guys know, right?
[08:33] He's super big in AI space,
[08:34] he does a lot of research,
[08:35] he pushed the limits of AI engineering
[08:39] when LAM was new, right?
[08:41] So he's a very big figure,
[08:43] you can go Google him.
[08:44] But the sweet spot that he found
[08:46] is do not go past 40 to 60% of your context window.
[08:50] That's where you start to see context rot,
[08:52] context drift and stuff like that, right?
[08:54] So 46% sweet spot.
[08:57] So what are the goals of frequent intentional compaction?
[09:01] It is to reduce context drift,
[09:03] reduce context rot and is to improve the output quality.
[09:06] And it is definitely not to save tokens, right?
[09:09] So if you are using these just to save tokens,
[09:11] maybe there might be a better way.
[09:13] But what we know for sure is,
[09:14] there's definitely different ways to save token.
[09:17] This is not a goal of intentional compaction, right?
[09:20] We just want to improve our output quality.
[09:24] So that is for frequent intentional compaction.
[09:26] The next one is research, plan and implement.
[09:29] It is pretty straightforward,
[09:31] it's a three stage process, right?
[09:33] And it's kind of self-explanatory.
[09:36] So it's three different stage
[09:37] and most of the work is actually being done
[09:40] on the research and the planning stage, right?
[09:43] So I would say even up to 85% of the time
[09:45] you should be spending on is the research
[09:47] and planning stage.
[09:49] Implementation, it should be as simple
[09:51] as just hitting generate, right?
[09:54] So let's take a look at what you look like.
[09:56] And, you know, RPI, this is designed
[09:59] to one shot a very huge task accurately, right?
[10:02] And let's assuming if you've done the research
[10:03] and the planning stage, right?
[10:06] So when we are talking about research, what do we do?
[10:08] We want to understand the system, data flows,
[10:10] architectures, stack, finding relevant files,
[10:13] configuration, maybe your CI configuration,
[10:16] how do you deploy, you know,
[10:17] what kind of Kubernetes environment you need to take note of
[10:20] and all these kinds of contexts,
[10:22] all these are factual contexts, you know,
[10:24] when we are doing a research,
[10:25] we want to find out factual stuff about a code base.
[10:27] We do not want to hunt for bugs,
[10:30] find improvements, find flaws, right?
[10:32] This is not the goal of researching.
[10:34] Researching is just to state all the facts
[10:36] about your code base, right?
[10:38] And this is the first step.
[10:40] Maybe your output can be something
[10:41] like research00.markdown, right?
[10:45] And then we move on to plan.
[10:46] So why is planning?
[10:47] Basically, the goal is to find and outline
[10:50] all your suitable implementation.
[10:53] You could include important file names, code snippets,
[10:56] and you could even be explicit,
[10:58] very explicit about your testing steps.
[11:00] What is your definition of done?
[11:02] What do you expect after this implementation
[11:05] is considered to be successful, right?
[11:07] So this is a planning stage.
[11:08] You could output it as, you know,
[11:10] plan feature x.markdown.
[11:12] So far, we have two markdown files, two outputs, right?
[11:17] Research and planning.
[11:18] And assuming if you do it right, implementing is simple.
[11:22] You just say refer to the first markdown file
[11:25] and implement, you know, let's say planning,
[11:27] loginintegration.markdown.
[11:31] And that's it.
[11:32] Implementation is supposed to be super quick.
[11:35] It's supposed to be like short
[11:38] after you have done your research and your planning, right?
[11:41] And that's all you gotta do to implement.
[11:43] So it's a three-step process.
[11:45] It is a workflow and there's no hard rules, right?
[11:48] It's not that there's some magic in a prompt.
[11:52] There's no magic in a prompt.
[11:53] It's just your process.
[11:55] And this RPI allows you to stay in control
[11:58] of your architecture and your decisions, right?
[12:00] You're not going through an agentic loop
[12:02] and finding out only when something hits production, right?
[12:06] So the goal of RPI is to be in control
[12:09] of what you're deciding, you know?
[12:10] It does not follow any specific prompts.
[12:12] It's meant to be generic.
[12:14] Sometimes, just sometimes,
[12:16] you feel that you don't have to do more research, right?
[12:19] Once you've got your research down,
[12:20] maybe you're doing a lot of planning and implementation,
[12:24] planning, implementation, right?
[12:25] You do not always need to do research,
[12:27] but you know, you don't have to follow anything specific.
[12:30] This is just a guideline, right?
[12:34] So some reminders before we quickly close off,
[12:41] AI does not read our mind yet, at least, at least, right?
[12:45] Who knows what some same old man is gonna come out tomorrow.
[12:48] But AI doesn't read our mind yet.
[12:50] So please, please do not, you know,
[12:53] for those of you that maybe are new,
[12:55] do not come up with prompts like this
[12:57] and expect very good results, right?
[12:59] They cannot read our mind.
[13:00] So try to be as accurate as possible,
[13:02] feed it as much correct context as possible.
[13:05] Don't just say, fix it, do your best, make no mistakes,
[13:09] right?
[13:10] And yeah, that's basically, you know,
[13:12] the goal, two methods for you all
[13:15] to improve your output, reduce slope.
[13:18] I have some references here,
[13:21] but don't worry, I'm gonna send it into a Discord group.
[13:24] In our Swiss Discord group and you guys can check it out.
[13:28] I've followed all these references and I got this idea.
[13:31] That's why I started this topic, right?
[13:34] But before we wrap up, maybe some of you are asking,
[13:37] so what happens to the sweet portion?
[13:40] Remember, the first slide says,
[13:42] agentic development is sweet, right?
[13:44] So where is the sweet portion?
[13:46] This is the sweet portion, right?
[13:48] So if you guys scan the QR code,
[13:49] you guys will be brought to Swiss MCP,
[13:52] official MCP server,
[13:54] and as simple as it sounds,
[13:57] this is all you need to get started, right?
[14:00] We have talked about, you know, how do you prom?
[14:02] How do you get a good process?
[14:04] How do you reduce hallucination?
[14:06] How do you improve your output?
[14:08] Using Swiss MCP server will further reduce
[14:12] your hallucinations, right?
[14:14] You don't even need to know what tech stack do we have.
[14:18] Just connect your agent to the SUE MCP and you'll find out.
[14:21] It's connected to our official most up-to-date docs.
[14:24] You can find everything there.
[14:25] You can trust that if your agent is retrieving information
[14:29] from there, you should be good, right?
[14:32] And I'm not just gonna say,
[14:33] I'm just gonna do a very quick demo.
[14:35] Do we have time?
[14:36] We have plenty of time actually.
[14:38] So I'm just gonna do a very quick demo
[14:39] to show you that it should work, right?
[14:43] So I have a cloud session here,
[14:46] and I basically just asked, are you connected, right?
[14:48] So it is connected.
[14:50] So let's just try it out.
[14:51] Let's just try what difference does it make
[14:53] when we have a sweetox MCP connected.
[15:00] Does it deploy contracts on the move?
[15:03] It doesn't automatically, but it will set up your skills.
[15:06] Yes.
[15:07] I want to create something,
[15:12] create something in place,
[15:16] what are some of my inputs.
[15:31] So without even knowing about tech stack,
[15:34] right, I begin with the SweetDogs MCP,
[15:37] and that's basically, if you have any idea,
[15:41] any, anything you want to build for hackathons,
[15:45] as long as you want to do it on Sweet,
[15:47] just hook it up to MCP,
[15:49] let your agent do the work,
[15:51] and trust that it's going to hallucinate way lesser.
[15:54] Right.
[15:59] It gives you some tips, decide what you're selling.
[16:02] Number two, kiosk basics.
[16:04] What is a kiosk?
[16:06] You don't know, but now you know that they're searching as a kiosk.
[16:09] Right.
[16:10] In Sweet, we have a kiosk,
[16:11] and kiosk is used to,
[16:12] is a primitive that is used to handle all kinds of marketplace integrations.
[16:17] Right.
[16:18] It's going to tell you what shared versus own objects.
[16:21] You don't even need to know what is that.
[16:23] You don't know what you don't know, right,
[16:24] but a dog is going to tell you what you need to take note of.
[16:27] So this is what I mean by, you know,
[16:29] just hook it up to your MCP,
[16:31] and you'll find everything related to your idea that you need to start building on.
[16:35] So I think it's super simple,
[16:38] and it's super quick to get started,
[16:40] and I highly recommend everyone to basically just get started.
[16:45] If you're new to Sweet, sure.
[16:46] Even if you're not new to Sweet,
[16:48] you could be using Sweet Dogs to find out more.
[16:50] And yeah, hopefully you guys can integrate and make something come true.
[16:57] Can I have a show of hands, roughly,
[16:59] who is going to build on Sweet today?
[17:03] Okay.
[17:04] Almost everyone.
[17:05] That's good.
[17:06] That's good, right?
[17:07] So if you guys already know what you guys are building,
[17:08] what stack are you using, you know,
[17:10] what are the primitives that you're using to make your project come true?
[17:14] Sure.
[17:15] If you don't know, hook it up and just ask away.
[17:18] But yeah, that's all for my presentation today.
[17:22] Five more minutes.
[17:24] Three more minutes, actually.
[17:25] Five more minutes.
[17:27] I'm open to any questions if you guys have any.
[17:30] I'll try my best.
[17:31] But if not, thank you.