From 36% to 100%: How Self-Improving Agents Write Their Own Skills — Rafal Wilinski, Runlayer

Hello everyone. I am Hello everyone. I am Rafael. Today I Rafael. Today I Rafael. Today I want to talk about want to talk about self-improving agents, self-improving agents, but not in a but not in a but not in a single-player mode where single-player mode where single-player mode where only only only one agent gets smarter. I want to one agent gets smarter. I want to one agent gets smarter. I want to discuss how to discuss how to discuss how to apply this apply this apply this concept, concept, concept, combined with combined with combined with network effects, network effects, network effects, to create a to create a to create a company or even an company or even an company or even an enterprise that enterprise that enterprise that self-improves self-improves self-improves while you sleep. Imagine the while you sleep. Imagine the while you sleep. Imagine the best engineer best engineer best engineer you've ever you've ever you've ever worked with, or worked with, or worked with, or salesperson, or lawyer. salesperson, or lawyer. The one who, every time he The one who, every time he solved a really solved a really solved a really difficult problem, difficult problem, difficult problem, wrote down exactly how wrote down exactly how wrote down exactly how he did it and he did it and he did it and passed the instructions on to passed the instructions on to passed the instructions on to the team. So when the team. So when the team. So when they encounter a they encounter a they encounter a problem like this, problem like this, problem like this, they don't need to they don't need to they don't need to solve it from solve it from solve it from scratch, right? scratch, right? scratch, right? Now imagine this Now imagine this Now imagine this happening with happening with happening with every complex every complex every complex problem across problem across problem across your entire company, and your entire company, and your entire company, and automatically every time. automatically every time. This is a dream, right? But This is a dream, right? But it almost never it almost never it almost never works that way. works that way. Even when you Even when you solve a very solve a very solve a very complex problem, complex problem, complex problem, often the knowledge, all the often the knowledge, all the often the knowledge, all the tuning, and all the tuning, and all the tuning, and all the reasoning just reasoning just reasoning just disappears. Oh, sorry.

disappears. Oh, sorry. All of these considerations All of these considerations simply disappear when simply disappear when simply disappear when you move on to the you move on to the you move on to the next task next task next task or close the session. or close the session. But I want to convince But I want to convince you that this is actually you that this is actually you that this is actually possible and we already possible and we already possible and we already have all the necessary have all the necessary have all the necessary elements for this. elements for this. If this sounds If this sounds interesting, let me interesting, let me interesting, let me introduce myself again. introduce myself again. introduce myself again. I am Rafael. I am a I am Rafael. I am a founding engineer at Runlayer. We founding engineer at Runlayer. We founding engineer at Runlayer. We create a "golden create a "golden create a "golden path" for path" for path" for implementing AI in implementing AI in implementing AI in your company. Before your company. Before your company. Before that, I managed Zapier that, I managed Zapier that, I managed Zapier agents and various agents and various agents and various AI teams, but in AI teams, but in AI teams, but in general, my general, my general, my previous work previous work previous work was about was about was about making agents not making agents not making agents not just flashy just flashy just flashy demos, but demos, but demos, but reliable reliable reliable tools in tools in tools in production. Before production. Before production. Before we get to the we get to the we get to the interesting parts, I interesting parts, I interesting parts, I want to break my want to break my want to break my topic into three separate topic into three separate topic into three separate parts and start parts and start parts and start simply with improving simply with improving simply with improving agents. There are agents. There are agents. There are many ways to do this. We many ways to do this. We many ways to do this. We can can can use use use better models, better better models, better better models, better tools for tools for tools for prompts, prompts, prompts, environments, etc., but environments, etc., but environments, etc., but today I want to today I want to today I want to focus only on focus only on focus only on skills because they are skills because they are skills because they are very accessible, very accessible, very accessible, right? At least in right? At least in right? At least in theory. To theory. To theory. To make sure we make sure we make sure we understand each understand each understand each other, for me other, for me other, for me a skill is an a skill is an a skill is an instruction that an agent instruction that an agent instruction that an agent can read when they can read when they can read when they feel that knowledge is feel that knowledge is feel that knowledge is relevant.

relevant. relevant. Technically, Technically, Technically, this term this term this term is called is called is called progressive progressive progressive disclosure, and it disclosure, and it disclosure, and it means that at the very means that at the very means that at the very beginning, the system or beginning, the system or beginning, the system or client only sees a client only sees a client only sees a brief summary or brief summary or brief summary or description of the skill. But description of the skill. But description of the skill. But when an agent understands when an agent understands when an agent understands that the knowledge inside a that the knowledge inside a that the knowledge inside a skill can be skill can be skill can be useful, it can useful, it can useful, it can read it in its read it in its read it in its entirety, add it to the entirety, add it to the entirety, add it to the main context, main context, main context, and and and change the trajectory of change the trajectory of change the trajectory of its actions based on this knowledge. Why is this so its actions based on this knowledge. Why is this so its actions based on this knowledge. Why is this so important? Well, unless you've important? Well, unless you've important? Well, unless you've been living in isolation for been living in isolation for been living in isolation for the last 6–12 months, the last 6–12 months, the last 6–12 months, your workload your workload has probably changed significantly, and has probably changed significantly, and that's because agents that's because agents that's because agents have gotten a lot have gotten a lot have gotten a lot better, or rather— better, or rather— models have gotten a lot models have gotten a lot models have gotten a lot better. And it's not better. And it's not better. And it's not just a feeling, it just a feeling, it just a feeling, it can be measured. Here can be measured. Here can be measured. Here you see a graph you see a graph you see a graph that you've probably that you've probably that you've probably seen too seen too seen too many times. For those many times. For those many times. For those who don't know, this is the who don't know, this is the who don't know, this is the MEETER benchmark. It MEETER benchmark. It MEETER benchmark. It measures how long measures how long measures how long agents can agents can agents can work independently work independently while still maintaining a while still maintaining a while still maintaining a decent chance of decent chance of decent chance of completing the task. And completing the task. And completing the task. And the word "chance" is the word "chance" is the word "chance" is key here. We'll key here. We'll key here. We'll come back to this in a come back to this in a come back to this in a moment, but moment, but moment, but the trend is clear.

the trend is clear. the trend is clear. Advanced capabilities, Advanced capabilities, Advanced capabilities, advanced intelligence advanced intelligence advanced intelligence are getting better and better. are getting better and better. are getting better and better. They They They do better with do better with do better with multi-step multi-step multi-step deep work. There is deep work. There is deep work. There is only one only one only one caveat: unless caveat: unless caveat: unless this advanced this advanced this advanced intelligence is intelligence is intelligence is taken away from you. And I am taken away from you. And I am taken away from you. And I am as prone to as prone to as prone to mistakes as mistakes as mistakes as you probably are. But there is a nuance. With you probably are. But there is a nuance. With you probably are. But there is a nuance. With great power or great power or great power or long-term long-term long-term tasks comes tasks comes tasks comes great great great responsibility. responsibility. responsibility. If an agent can If an agent can If an agent can reason for much reason for much reason for much longer, it also longer, it also longer, it also means that it can means that it can means that it can go down a very, go down a very, go down a very, very deep rabbit very deep rabbit very deep rabbit hole. He can hole. He can hole. He can spend hundreds of spend hundreds of tool calls tool calls trying to trying to trying to defend a false defend a false defend a false thesis, argue thesis, argue thesis, argue for hours, and the next for hours, and the next for hours, and the next agent will do exactly the agent will do exactly the agent will do exactly the same thing, or perhaps same thing, or perhaps same thing, or perhaps even worse. even worse. even worse. Therefore, the initial Therefore, the initial Therefore, the initial trajectory is of trajectory is of trajectory is of great importance. And that's why great importance. And that's why great importance. And that's why skills are skills are skills are so important, because so important, because so important, because they direct the they direct the they direct the agent to the right agent to the right agent to the right goal. Otherwise, you will goal. Otherwise, you will goal. Otherwise, you will waste time, waste time, waste time, waste tokens, and waste tokens, and waste tokens, and most importantly, most importantly, most importantly, waste money. waste money. Valuable skills are those Valuable skills are those rituals that no one rituals that no one rituals that no one wants to wants to wants to rediscover. This knowledge is rediscover. This knowledge is rediscover. This knowledge is often not something often not something often not something exciting. They are exciting. They are exciting. They are probably deeply probably deeply probably deeply ingrained in your ingrained in your ingrained in your company. And what if company. And what if company. And what if agents were trying agents were trying agents were trying to discover this knowledge from to discover this knowledge from to discover this knowledge from scratch—for example, scratch—for example, scratch—for example, the sequence of the sequence of the sequence of scripts and scripts and scripts and function switches function switches function switches to create a good to create a good to create a good roadmap, or the roadmap, or the roadmap, or the exact wording exact wording exact wording to communicate with that to communicate with that to communicate with that picky picky picky corporate corporate corporate client who is about to leave client who is about to leave you. Yes, you. Yes, you. Yes, if agents if agents if agents tried to tried to tried to figure this out from scratch, figure this out from scratch, figure this out from scratch, they would most likely they would most likely they would most likely fail. And

fail. And fail. And often this failure can often this failure can often this failure can be very costly. be very costly. be very costly. This is why I believe This is why I believe This is why I believe so many so many so many AI implementations or AI implementations or AI implementations or pilots pilots pilots fail: you don't fail: you don't fail: you don't provide agents with the provide agents with the provide agents with the appropriate appropriate appropriate context and context and context and instructions on instructions on instructions on how to act within your how to act within your how to act within your company. So, company. So, company. So, skills are great, skills are great, skills are great, but I think but I think but I think they're not without their they're not without their they're not without their problems. There are three problems. There are three problems. There are three major major major issues, and the first one is that issues, and the first one is that they are local and developer-centric. What I developer-centric. What I mean is: when you mean is: when you mean is: when you look at the look at the look at the skills ecosystem, skills ecosystem, skills ecosystem, you'll find markdown you'll find markdown files, JSON files, CLI files, JSON files, CLI commands, and commands, and commands, and Git repositories. And if Git repositories. And if Git repositories. And if you're a developer, that's you're a developer, that's you're a developer, that's great. But if you’re a great. But if you’re a great. But if you’re a VP of VP of VP of AI or an AI or an AI implementation lead, you AI implementation lead, you need to be more need to be more need to be more empathetic about empathetic about empathetic about how to roll out an how to roll out an how to roll out an AI strategy across the AI strategy across the AI strategy across the company, including company, including company, including non-technical non-technical non-technical people, right? people, right? And the main challenge here is And the main challenge here is how non-technical how non-technical how non-technical people can people can people can create and create and create and distribute these distribute these distribute these skills. Oh, I love skills. Oh, I love skills. Oh, I love this meme. This is a junior this meme. This is a junior developer or developer or developer or non-technical professional non-technical professional who discovers the who discovers the who discovers the NPX skills command— NPX skills command— the most important and the most important and the most important and popular popular popular command for command for command for installing them. This is the installing them. This is the installing them. This is the only command I know only command I know only command I know that that that installs not just a installs not just a installs not just a skill, but also skill, but also skill, but also free free free prompt injections, a wild prompt injections, a wild prompt injections, a wild script with root privileges, script with root privileges, script with root privileges, and someone else's API key—all and someone else's API key—all and someone else's API key—all in one run.

in one run. Believe me, at RunLayer we Believe me, at RunLayer we scan scan scan thousands of skills every day, and some thousands of skills every day, and some thousands of skills every day, and some really really really weird things happen there. Okay, the weird things happen there. Okay, the weird things happen there. Okay, the second problem: I think second problem: I think we all we all we all like the situation like the situation when we are faced when we are faced when we are faced with a difficult with a difficult with a difficult problem and we already have a problem and we already have a problem and we already have a ready-made plan of action on how to ready-made plan of action on how to ready-made plan of action on how to solve it, right? But solve it, right? But solve it, right? But often there is often there is often there is simply no such plan. Because it simply no such plan. Because it simply no such plan. Because it takes takes takes time, effort, and time, effort, and time, effort, and energy to energy to energy to understand the problem and understand the problem and understand the problem and leave something leave something leave something useful for posterity. useful for posterity. And the third problem is that And the third problem is that not every LLM not every LLM client supports client supports client supports the skills, right? If the skills, right? If the skills, right? If you're in the cloud you're in the cloud you're in the cloud ecosystem, then ecosystem, then ecosystem, then everything is probably great. everything is probably great. everything is probably great. If you're a developer, you'll likely be If you're a developer, you'll likely be fine, even fine, even if the if the if the skills support is weird, skills support is weird, skills support is weird, depending on depending on depending on which client you're which client you're which client you're using. And using. And using. And again, if again, if again, if you think about marketing, HR, you think about marketing, HR, you think about marketing, HR, legal, finance legal, finance —these departments are —these departments are —these departments are probably still probably still probably still using ChatGPT using ChatGPT using ChatGPT or some other LLM or some other LLM client where the skills client where the skills client where the skills may not be may not be may not be supported. There supported. There was a big was a big heated discussion a while ago when heated discussion a while ago when heated discussion a while ago when these skills first these skills first these skills first emerged. Some emerged. Some emerged. Some have proclaimed: thank have proclaimed: thank have proclaimed: thank god MCP is finally god MCP is finally god MCP is finally dead, because I don't want to dead, because I don't want to dead, because I don't want to mess with OAuth. Now mess with OAuth. Now mess with OAuth. Now I can "hardcode I can "hardcode " a bunch of CLI commands, " a bunch of CLI commands, " a bunch of CLI commands, write my API key write my API key write my API key right into a markdown file, and right into a markdown file, and right into a markdown file, and call it progress, call it progress, call it progress, right? But, in my opinion, right? But, in my opinion, right? But, in my opinion, everyone missed the point.

everyone missed the point. everyone missed the point. Skills can Skills can Skills can convey knowledge convey knowledge convey knowledge about how to do something about how to do something , while tools , while tools , while tools give us give us give us capabilities. capabilities. capabilities. The tools with MCP The tools with MCP The tools with MCP give us give us give us the ability to do the ability to do the ability to do these things, but MCP can these things, but MCP can these things, but MCP can also be used as a also be used as a also be used as a protocol for protocol for protocol for transferring knowledge, for transferring knowledge, for transferring knowledge, for transferring skills. We transferring skills. We transferring skills. We can can can use MCP as a use MCP as a distribution layer. So, distribution layer. So, instead of installing instead of installing instead of installing skills locally, you skills locally, you skills locally, you can create a can create a can create a remote MCP server remote MCP server remote MCP server that will become the source of that will become the source of that will become the source of truth for truth for truth for your entire company. your entire company. your entire company. Clients can Clients can Clients can connect to connect to connect to it and search for it and search for it and search for the skills they need. the skills they need. the skills they need. The server can The server can enforce policies and enforce policies and ensure that everyone ensure that everyone ensure that everyone gets the best gets the best gets the best version of the skill version of the skill version of the skill based on the request and based on the request and , for example, the , for example, the , for example, the assigned role. We've been assigned role. We've been assigned role. We've been doing this internally doing this internally doing this internally at Run:ai for the last 6 at Run:ai for the last 6 at Run:ai for the last 6 months and it's working months and it's working months and it's working pretty well. But this is pretty well. But this is pretty well. But this is no longer just an no longer just an no longer just an internal matter. We internal matter. We internal matter. We see the broader see the broader see the broader MCP ecosystem MCP ecosystem MCP ecosystem moving in this moving in this moving in this direction as well. There is a direction as well. There is a direction as well. There is a working group working group working group called “Skills over MCP,” and called “Skills over MCP,” and called “Skills over MCP,” and while the exact API is not yet while the exact API is not yet while the exact API is not yet final, the direction is final, the direction is final, the direction is clear. MCP can be clear. MCP can be clear. MCP can be used to used to used to provide skills and provide skills and provide skills and spread knowledge throughout spread knowledge throughout spread knowledge throughout your company.

your company. Yes, and although the official Yes, and although the official specification specification specification leans towards leans towards leans towards using using using MCP resources, MCP resources, MCP resources, resource support is resource support is resource support is still patchy. still patchy. still patchy. Therefore, we focused on Therefore, we focused on Therefore, we focused on tools. And tools. And tools. And if you remember the if you remember the if you remember the first and third first and third first and third problems I problems I problems I mentioned— mentioned— developer focus and lack of developer focus and lack of customer support— customer support— then MCP as a distribution layer then MCP as a distribution layer solves that because solves that because it is unified and it is unified and it is unified and works in almost works in almost works in almost every client, not every client, not every client, not just development ones. just development ones. So the problem of So the problem of skills dissemination skills dissemination skills dissemination among non-technical among non-technical among non-technical professionals is essentially professionals is essentially professionals is essentially solved, right? solved, right? solved, right? Additionally, you now Additionally, you now Additionally, you now have a centralized API. have a centralized API. have a centralized API. So as a VP So as a VP So as a VP of AI or of AI or AI implementation leader, you AI implementation leader, you have a single place to have a single place to have a single place to manage and manage and manage and control what control what control what knowledge is shared across knowledge is shared across knowledge is shared across your company. your company. Okay, now we Okay, now we know how to improve know how to improve know how to improve agents because we have the agents because we have the agents because we have the skills. We know we skills. We know we skills. We know we can can can use MCP use MCP use MCP to spread to spread to spread these skills to these skills to these skills to all clients and all clients and all clients and agents, but let’s agents, but let’s agents, but let’s talk about something more talk about something more talk about something more interesting: how agents interesting: how agents interesting: how agents can can can improve improve improve themselves. And every time I themselves. And every time I themselves. And every time I see a new article or see a new article or see a new article or framework on the topic, framework on the topic, framework on the topic, I always go back I always go back I always go back to the work on Voyager.

to the work on Voyager. to the work on Voyager. Unfortunately, you can't see Unfortunately, you can't see Unfortunately, you can't see the video, but there is a Minecraft game playing in the background the video, but there is a Minecraft game playing in the background the video, but there is a Minecraft game playing in the background . In 2023, . In 2023, . In 2023, right after the release of right after the release of right after the release of GPT-4, that is, a long time GPT-4, that is, a long time ago, ago, ago, the question arose: how long the question arose: how long can multiple agents or systems can multiple agents or systems explore the explore the explore the Minecraft space? They placed Minecraft space? They placed Minecraft space? They placed the agent in Minecraft without the agent in Minecraft without the agent in Minecraft without any instructions. any instructions. any instructions. He set goals for himself He set goals for himself He set goals for himself . It . It . It just kept just kept just kept getting better getting better getting better without even changing the without even changing the without even changing the model weights or model weights or model weights or prompts. And in 2023, I prompts. And in 2023, I prompts. And in 2023, I think it was a think it was a think it was a real breakthrough. real breakthrough. real breakthrough. How did they do it? How did they do it? The main thing to pay attention to here is the is the skills library. So Voyager didn't skills library. So Voyager didn't skills library. So Voyager didn't just thoroughly just thoroughly just thoroughly explore explore explore Minecraft space and then Minecraft space and then Minecraft space and then discard the discard the discard the data it received. When an agent data it received. When an agent data it received. When an agent discovered something new, discovered something new, discovered something new, it stored a recipe in the it stored a recipe in the it stored a recipe in the form of code form of code form of code consisting of calls to the consisting of calls to the consisting of calls to the tools that tools that tools that led it to the led it to the led it to the discovery. So discovery. So discovery. So later, when the agent later, when the agent later, when the agent decided, "Hey, I decided, "Hey, I decided, "Hey, I need to create a need to create a need to create a workbench or a diamond workbench or a diamond workbench or a diamond pickaxe." He didn't pickaxe." He didn't pickaxe." He didn't have to have to have to invent all this invent all this invent all this knowledge from scratch. He could knowledge from scratch. He could knowledge from scratch. He could simply go to the simply go to the simply go to the skill library and skill library and skill library and get it as code get it as code —as a recipe that —as a recipe that —as a recipe that could be could be could be reused. And this is reused. And this is reused. And this is actually very actually very actually very important. The main idea important. The main idea I want to convey is I want to convey is I want to convey is that that that research should be research should be research should be transformed into transformed into reusable procedural knowledge. And we reusable procedural knowledge. And we started doing something started doing something started doing something very similar in Run Layer. We very similar in Run Layer. We very similar in Run Layer. We let the agents let the agents let the agents do the real

do the real do the real work, and when the run is work, and when the run is work, and when the run is successful, successful, successful, or when there's something interesting in it— or when there's something interesting in it— like a few tool calls—we tool calls—we just take that just take that just take that trace and ask Frontier LLM trace and ask Frontier LLM trace and ask Frontier LLM to highlight a skill. to highlight a skill. While it's extremely easy to isolate While it's extremely easy to isolate a skill from a single a skill from a single a skill from a single run run run , I believe that , I believe that , I believe that we, like agents, we, like agents, we, like agents, learn from mistakes. learn from mistakes. True True improvement improvement improvement occurs precisely occurs precisely occurs precisely when an agent when an agent when an agent fails. So, a fails. So, a fails. So, a launch that didn't go launch that didn't go launch that didn't go as planned as planned as planned is actually the richest is actually the richest is actually the richest source of information, source of information, source of information, because it reveals because it reveals because it reveals missing missing missing fuses, fuses, fuses, libraries, or libraries, or libraries, or some edge some edge some edge cases that you cases that you cases that you didn't think about didn't think about didn't think about beforehand. And each beforehand. And each beforehand. And each launch, successful or launch, successful or launch, successful or not, fuels our not, fuels our not, fuels our distillation mechanism, distillation mechanism, distillation mechanism, constantly constantly constantly refining our refining our refining our skills. Successes skills. Successes skills. Successes make skills make skills make skills more reliable, because more reliable, because more reliable, because if the agent achieved the goal if the agent achieved the goal , it means "Hey, , it means "Hey, , it means "Hey, this skill works." And this skill works." And this skill works." And if not, it makes if not, it makes if not, it makes the skill more the skill more the skill more sustainable, because now we sustainable, because now we sustainable, because now we know what exactly was know what exactly was know what exactly was missing here. And all this missing here. And all this missing here. And all this happens happens happens completely autonomously completely autonomously completely autonomously and asynchronously, without and asynchronously, without and asynchronously, without human intervention. The human intervention. The initial results initial results were excellent. We were were excellent. We were were excellent. We were tasked with tasked with tasked with running a Chromium fork running a Chromium fork running a Chromium fork inside AWS Lambda inside AWS Lambda inside AWS Lambda using Sonnet 4.6 using Sonnet 4.6 . It's not the . It's not the . It's not the most intelligent most intelligent most intelligent model, but model, but model, but it has sometimes succeeded.

it has sometimes succeeded. it has sometimes succeeded. However, the However, the However, the 36% success rate was not ready for 36% success rate was not ready for 36% success rate was not ready for industrial industrial industrial use. So we use. So we use. So we enabled the agent enabled the agent enabled the agent self-improvement feature self-improvement feature self-improvement feature , and during , and during , and during one of the runs, one of the runs, one of the runs, the agent found a the agent found a the agent found a solution, and our solution, and our solution, and our skill distiller skill distiller skill distiller created a created a created a reusable reusable reusable algorithm. It was algorithm. It was algorithm. It was included in all included in all included in all future launches and the future launches and the future launches and the success rate increased to success rate increased to success rate increased to 100%. So yeah, that's great. 100%. So yeah, that's great. 100%. So yeah, that's great. It's like a developer being It's like a developer being It's like a developer being more concerned with more concerned with more concerned with documentation. Um, documentation. Um, documentation. Um, success rate is success rate is success rate is the easiest to show, the easiest to show, the easiest to show, but a good skill but a good skill is narrowing the is narrowing the is narrowing the search space. This forces the search space. This forces the search space. This forces the agent to Gm, it narrows the agent to Gm, it narrows the agent to Gm, it narrows the search space, and the search space, and the search space, and the agent stops agent stops agent stops making making making tool calls, tool calls, tool calls, exploring fewer exploring fewer exploring fewer branches. This saves branches. This saves branches. This saves tokens and reduces tokens and reduces tokens and reduces costs. But, costs. But, costs. But, putting aside economy putting aside economy putting aside economy and precision, I think and precision, I think and precision, I think there's a deeper reason there's a deeper reason there's a deeper reason why skills are worth why skills are worth why skills are worth singling out singling out singling out . Um, the problem is that . Um, the problem is that . Um, the problem is that advanced advanced advanced intelligence is a intelligence is a intelligence is a rented or rented or rented or borrowed resource. He borrowed resource. He borrowed resource. He may disappear. I think, may disappear. I think, may disappear. I think, yes, we all miss yes, we all miss yes, we all miss Fable. It was a great Fable. It was a great Fable. It was a great model, wasn't it? It model, wasn't it? It model, wasn't it? It may become outdated or may become outdated or may become outdated or it may become it may become it may become stupider. We complain every day stupider. We complain every day stupider. We complain every day on Twitter or on Twitter or on Twitter or Slack that Opus is Slack that Opus is Slack that Opus is really stupid these days, right?

really stupid these days, right? really stupid these days, right? So while we have So while we have So while we have access to the best access to the best access to the best version of this version of this version of this intelligence, we should intelligence, we should intelligence, we should use it use it use it to discover and to discover and to discover and preserve useful preserve useful preserve useful routines. We should routines. We should routines. We should create a catalog of create a catalog of create a catalog of how how how work is done in our work is done in our work is done in our company, as a company, as a company, as a collective collective collective subconscious. And how do subconscious. And how do subconscious. And how do we do it? Well, I we do it? Well, I we do it? Well, I believe believe believe we already have all the pieces of the puzzle. Um, we already have all the pieces of the puzzle. Um, we already have all the pieces of the puzzle. Um, we have skills as a we have skills as a we have skills as a way to preserve way to preserve way to preserve procedures, procedures, procedures, improve improve improve agents, and manage agents, and manage agents, and manage models. We have MCP as the models. We have MCP as the models. We have MCP as the only only only distribution level for distribution level for distribution level for these skills. And these skills. And these skills. And we also have a central we also have a central we also have a central repository, a knowledge base repository, a knowledge base repository, a knowledge base for all these skills, for all these skills, for all these skills, which is constantly which is constantly which is constantly updated and updated and updated and supplemented with new supplemented with new supplemented with new data. Um, yes, we data. Um, yes, we data. Um, yes, we also have a also have a also have a continuous continuous continuous learning flywheel for a single learning flywheel for a single learning flywheel for a single agent, but what's interesting is agent, but what's interesting is agent, but what's interesting is what I promised you at the what I promised you at the what I promised you at the very beginning—the very beginning—the network effect. Um, network effect. Um, network effect. Um, so what happens if you so what happens if you so what happens if you multiply that by multiply that by multiply that by every agent and every agent and every agent and every customer in the every customer in the every customer in the entire company. At Runway AI, entire company. At Runway AI, entire company. At Runway AI, we call this the we call this the we call this the organizational organizational organizational flywheel flywheel flywheel of self-improvement.

of self-improvement. of self-improvement. Um, we're grouping a number of Um, we're grouping a number of Um, we're grouping a number of similar launches and similar launches and similar launches and trying to trying to trying to isolate the skill. isolate the skill. isolate the skill. It goes through It goes through It goes through our numerous our numerous our numerous checks, like checks, like checks, like whether it contains any whether it contains any whether it contains any prompt injections or prompt injections or prompt injections or PII, um, and many PII, um, and many PII, um, and many other things. When it's other things. When it's other things. When it's ready, it ready, it ready, it goes into our goes into our goes into our skills library and, skills library and, skills library and, through our through our through our MCD gateway, is made available to every MCD gateway, is made available to every MCD gateway, is made available to every customer and agent in the customer and agent in the customer and agent in the company. As company. As company. As a result, we have a a result, we have a a result, we have a knowledge base that is knowledge base that is knowledge base that is constantly constantly constantly evolving, always evolving, always evolving, always relevant, and relevant, and relevant, and continuously continuously continuously updated with new updated with new updated with new information and information and information and new edge new edge new edge cases. So once cases. So once cases. So once one agent one agent one agent solves a very solves a very solves a very complex problem, complex problem, complex problem, it becomes a it becomes a it becomes a skill that skill that skill that is stored in the is stored in the is stored in the shared shared shared organizational organizational organizational library. So library. So library. So when someone else when someone else encounters the same encounters the same problem in the future, they won't have to problem in the future, they won't have to problem in the future, they won't have to find a find a find a solution from scratch. They don't solution from scratch. They don't solution from scratch. They don't necessarily need to know, uh necessarily need to know, uh , the exact sequence of , the exact sequence of , the exact sequence of those magic commands those magic commands those magic commands that will restart the that will restart the that will restart the database or database or database or roll back production. roll back production. roll back production. They can just, They can just, They can just, um, refer to the um, refer to the playbook they have. And why is this playbook they have. And why is this so important? Because I so important? Because I so important? Because I think 50% of us have think 50% of us have think 50% of us have probably already had an probably already had an probably already had an existential crisis existential crisis existential crisis because of these models, and because of these models, and because of these models, and the thoughts boil down to the thoughts boil down to the thoughts boil down to , uh, models , uh, models , uh, models becoming a commodity becoming a commodity becoming a commodity .

. . Everyone here has access to them Everyone here has access to them Everyone here has access to them . Um, the cost of . Um, the cost of . Um, the cost of software software is approaching zero. is approaching zero. Anyone can Anyone can Anyone can replicate my replicate my replicate my product, um, product, um, product, um, at least on the at least on the at least on the surface. In fact, surface. In fact, surface. In fact, so many people so many people so many people work at work at work at the show today and think, " the show today and think, " Hey, everyone's building the same thing Hey, everyone's building the same thing Hey, everyone's building the same thing , right?" , right?" , right?" So, what is my "moat"? So, what is my "moat"? Um, and I believe that this " moat" could be about moat" could be about moat" could be about creating this kind of creating this kind of creating this kind of flywheel and flywheel and flywheel and accumulating knowledge accumulating knowledge accumulating knowledge through it, because it's like through it, because it's like through it, because it's like your own your own your own hard-earned lessons, hard-earned lessons, hard-earned lessons, systematized and systematized and systematized and aligned with how aligned with how aligned with how your company works, your company works, your company works, you know? It's you know? It's you know? It's captured, it's captured, it's captured, it's accumulated, accumulated, accumulated, raising the bar raising the bar raising the bar for everyone in the for everyone in the for everyone in the company, giving them company, giving them company, giving them access to the best access to the best access to the best knowledge available. Um, knowledge available. Um, knowledge available. Um, it's also a way to become it's also a way to become it's also a way to become resistant to resistant to model intelligence obsolescence, and model intelligence obsolescence, and it's a great way it's a great way it's a great way to reduce costs to reduce costs to reduce costs without even changing the without even changing the without even changing the model. So, yes. This is model. So, yes. This is model. So, yes. This is perhaps the only thing a perhaps the only thing a perhaps the only thing a competitor will not be able to competitor will not be able to competitor will not be able to copy—your copy—your copy—your internal culture internal culture internal culture and identity. And and identity. And and identity. And yes, that's practically all.

yes, that's practically all. yes, that's practically all. Um, thanks for coming. Um, thanks for coming. Um, thanks for coming. If this resonated If this resonated If this resonated with you, I am Raphael. I with you, I am Raphael. I with you, I am Raphael. I work at Front Layer. We work at Front Layer. We work at Front Layer. We are developing for you the " are developing for you the " golden path" of golden path" of golden path" of using AI. I'm using AI. I'm using AI. I'm open to talking open to talking open to talking about skills, about skills, about skills, MCP agents, and more. And that's exactly what MCP agents, and more. And that's exactly what MCP agents, and more. And that's exactly what you build in Front you build in Front you build in Front Layer. Thank you.