Speaker A
Hello community. So great that you are back today. Today we're bringing everything together from last week.
Explores Google's new agentic skill framework as a dynamic, versioned procedural OS for AI, advancing skill definition and system architecture.
Key Takeaways
- Agentic skills represent a fundamental shift from static procedures to dynamic, modular state-transforming programs.
- Skills must meet strict criteria to ensure modularity, relevance, and effective state transformation.
- Skill management requires lifecycle and memory considerations to maintain system efficiency and reliability.
- The new framework supports complex multi-agent and swarm intelligence applications through dynamic skill invocation.
- Google's architecture integrates skill execution with verification and security, forming a procedural operating system for AI.
What the video covers
- The video reviews recent top AI papers, focusing on Google's 2026 paper on agentic skills as a system foundation.
- Agentic skills are redefined as versioned, permission-bounded, dynamically routed programs transforming agent or environmental states.
- Skills evolve from simple instructions to complex state transition operators with formal six-element tuples.
- Three tests define what qualifies as a skill: encapsulated modularity, late-bound invocation, and dual state transformation.
- Skills differ from procedures by encoding dynamic state changes rather than static instructions or workflows.
- Historical approaches to skill management are discussed, highlighting limitations of transcripts, hardcoded workflows, and tool-based methods.
- The new framework integrates skill lifecycle management, including discovery, acquisition, authoring, storage, retrieval, routing, and orchestration.
- Skill libraries are treated as memory management problems with retention and eviction policies based on skill competence metrics.
- The system emphasizes verification, security, and dynamic invocation within multi-agent or swarm intelligence contexts.
- Google's model connects procedural memory with execution, routing, and environmental feedback into a cohesive AI operating system.
Chapters
- 00:00Introduction and Overview
- 01:52Basic Overview of Procedural Skill Lifecycle
- 03:59New System View of Agentic Skills
- 06:21Tests Defining a Skill
- 08:16Challenges with Previous Skill Approaches
- 10:29Harness and System Transition Verification
- 12:44Skill Libraries and Memory Management
- 14:46Skill Competence and Retention Metrics
- 19:20Skill Discovery, Validation, and State Transition
- 23:58LLM Responsibilities and System Integration
Full Transcript — Download SRT & Markdown
Speaker A
Today we have the master class. So let's have a look. Agentic skills and now we have the complete systems layer architecture. Let's go for it. So let's start. What are here? September 1st, the top three papers here globally. And I
Speaker A
scanned about 200 here really manually. I read here the scientific abstract from 200 papers. And I think those are the top three of today. The first here, quantum physics and intelligence science, department of physics, Capital Normal University, Beijing. And they go,
Speaker A
LLMs interpret embeddings, organize and graph emerge. And they have an agent-driven compilation of scientific knowledge. A beautiful paper, August 30, 2026. Autoscientist quant self-evolving coding agents for automatic research in quantitative investments. You remember alphas ranked the stocks by expected
Speaker A
future returns under basic input for the quantitative investment calculations. A beautiful paper by Google. Yeah.
Speaker A
Published also here at September 1st. Remember do not look here for August 5th. It was published today. And then another paper here by this little company called Google. And they have towards a system foundation for agentic skills. The architecture, the lifestyle
Speaker A
and the security. And I said wait a minute, agentic skill, we have now a system foundation, what is it? And I decided okay skip here the financial paper and those are the two papers of today I'm going to
Speaker A
have a look at and we start here with the systems foundation for by Google.
Speaker A
So there are now multiple ways to read this paper. Now, the first way to read this paper is simple. You just go there and you say okay so we do have an agentic skill framework and if you're new to AI and you have not been with me
Speaker A
in the last 3 years on this channel you say okay here I get a basic overview here on procedural skill life cycle discovery, acquisition, authoring, representation, storage, memory retrieval, routing and orchestration and you have here all the most important paper on
Speaker A
this subject and then you go okay taxonomy of skills, core capabilities, domain specifics, software engineering, embodied robotics, ML skills, self-improvement, tool execution, AI, planning traces, reasoning traces. Here are the most important papers of 2026.
Speaker A
Great. So this is the first way to read this paper here as kind of a knowledge repo. But you know us, know here on this channel we go a level deeper. We want to have you the core idea. Now I think the
Speaker A
real story line of this paper is much stronger. Now Google is real clever in this presentation. No, it gives us a lot of historical data, a lot of literature, but they also give us here some new ideas. Looking at the literature,
Speaker A
looking at the complete development here on a system level, they say, hmm, is there something we can develop a little bit further and they say, you know, an agentic skill is not simply a good prompt or a stored procedure, a
Speaker A
workflow, how to do something, have a new capability for the system, no. And they say you know what a skill is now going here on the next level. A skill is now an agentic skill in particular is a
Speaker A
governed dynamically invoked state transition object inside the agent harness. And of course yeah mathematically we're going to construct the state transition operator of course but let's start here from the beginning.
Speaker A
So if you're new to AI, you know, at the very beginning, skills were just some useful instruction how to do something to achieve your particular sub-goal. And then skills evolved here to an external procedural memory. And suddenly we had
Speaker A
the bridge between skill markdown files and memory markdown files and their integration here into harness structures like cloud code or opex or codeex. So external protoal memory and a lot of you will be on this level. And now Google
Speaker A
tells us you know what we can go even further. And we see this now in a system view that a skill is a versioned permission-bounded dynamically routed program that transforms agent or environmental states and they came up with a new transition
Speaker A
here tuple with six elements. And they say a skill is now for Google here in September 2026 something new. It's a packaged state transformation tuple later becomes an operator. So what is this skill? So we have here quite a lot
Speaker A
of elements. Now A of course is activate when a dependency related build failure is detected. I our instruction that you know diagnosing resolving dependency conflict. C is guess what applicable only when the repo exposes compatible dependency. P is here our
Speaker A
file inspection on the package manager, the build system, the test runner. Pi is here our classical control flow and E are all the dependencies, resolution, successful verified build structures. So you see a skill is not anymore only some
Speaker A
instructional text. Here I screenshot from the paper. Beautiful. So they give you here the definition of a skill here as this six tuple. Then they define here the agent state. Then they define a skill activation. A skill activates
Speaker A
conditional based on the agent state x of t and the activation goal g where a here dependent here on x of t and g equal one indicates that the skill should trigger formalizing dynamic routing conceptual application and planner guided invocation and you got
Speaker A
it. So this means a skill is now if you want a complete contract not a complete complexity when is it relevant and under what valid conditions are we allowed to use it with what permitted tools is this executing this particular policy
Speaker A
and to produce your a particular set of effects. So it is not anymore just an instruction do this or have this workflow. It is now going here for complete complexity.
Speaker A
Now you remember in my last videos we were talking about trajectory and procedures here and a procedure is now according to this new idea by Google a procedure must now pass three particular tests to become now a skill to qualify
Speaker A
as a skill. The test one is encapsulated modularity can the skill be stored version updated and tested independently of the model weights and the global harness itself. Right. Test two late-bound invocation is skill loaded only when the current state and the goal
Speaker A
make it really relevant and test three is the approach to dual state transformation. Does the skill encode how to change the state rather than merely some stored facts or history and say hey wait a minute wait a minute
Speaker A
why why we have to do this? Well, a skill is not a procedure. No, a procedure disappeared here within the tons of text within the conversation that the AI is having or multiple AI systems are having. And the transcript,
Speaker A
yeah, of course, it records every word that happened. But a skill must now capture what should be done again. And it is a pattern recognition system that extracts now here a particular skill.
Speaker A
But you see a transcript and a skill are not identical objects and therefore we have these tests. Now you remember looking back in the past we had more and more sophisticated attempts that already guide us in this direction. Now the old
Speaker A
one saved the complete transcript and this is what we done in the old days.
Speaker A
Now too large too noisy here instance specific millions of token just append append append just absolute nonsense.
Speaker A
Then we said okay let's put the procedure now into the system prompt. Yeah it consumes context on return including all the irrelevant task. I said okay give them all the necessary tools no but tools provide capabilities but do not encode the procedure for
Speaker A
combining them for example. Yeah, another problem and then we said you know what let's hardcode now finally a workflow now let's give it this TDI system and the problem was yeah the workflow is persistent but fixed activation and fixed topology make it
Speaker A
difficult to use dynamically especially if we have a single agent multi-agent or swarm intelligence so you see yeah we had to move on think about this skill definition what it does the individual components may be implemented really
Speaker A
think But what is it? A the activation, maybe a rule, a classifier, some embedding router, or guess what? An LLM decision. Of course, either instructions that maybe in natural language, maybe in structured JSON, or in any formal specification that we have, pi, natural
Speaker A
language guidance or some executable code structure, a state machine or a deck, e textual expectation or mechanically verified post conditions.
Speaker A
Absolutely amazing. what degrees of freedom we have for the skill definition. But look at this. You immediately understand that this six tuple here of a skill s in this redefinition by Google is now an interface specification. Also an
Speaker A
interface specification and this expresses here the skill as a conditional state transition operator according to the mathematics Google presents to you in this paper. So what a skill does it a skill now is operating on a different level of understanding. A
Speaker A
skill attempts to move the complete system from a state x of t where we have an unresolved dependency failure to a next state x of t +1 where the dependencies are resolved and the build is verified. So we see a skill now in a
Speaker A
complete different system. Now think about this. What have we achieved? And also think back to the last four videos here on my channel.
Speaker A
Again the LLM the core of the agent may generate all the reasoning and all the action. It may decide on this beautiful with its neural network. A skill defines here the procedural not only the memory structure but here the procedural
Speaker A
transformation of the complete system. Now we need here an environment that provides here some feedback given our particular action that was decided here by the LLM and the environment produces here some detectable observation via our sensor structures. Beautiful. And then
Speaker A
we have the harness the deterministic harness and the harness now determines whether the intended transition of the complete system was actually achieved.
Speaker A
So you see even the positioning of your harness is now a system harness and not anymore a harness that has just a multitude of skills and a multitude of memory files that are more or less not interdependent.
Speaker A
Google also gives us here a classical introduction. How are skills created? No, they have their equation six where they define a discovery operator here as new. This is a set of unverified skill candidates that are simply here with D.
Speaker A
This is the discovery process here and dependent on h and h is a collection of execution trajectories simple task find here the reusable sub trajectories here all the pattern all the traces that we can extract here that we can reuse later
Speaker A
on. Yeah, discovery process looks for reusable sub trajectories that produce persistent environmental outcomes. here reduce the uncertainty or the future branching recur across the task and can be generalized. Yeah, absolutely. Beyond a single episode. Yeah. So here we have
Speaker A
it. You have now you select a skill formulation that maximizes the expected reward coincidence inside its applicable region for safety reason but emitted only if the validation condition passes.
Speaker A
So skill remember Wiki skill persistent skill compiler. Oh yeah, there's a lot of things to do. And guess what? This is also the transition here to the videos you already saw last week. Wiki skill and the dual memory system. Those videos
Speaker A
and those integrations are now compatible. And you might say, yeah, what a coincidence. No, Wiki skill is well was published by Google. So no wonder that we have now a coherent system approach after we have some components that were published earlier
Speaker A
by Google. So therefore, hm Google absolutely amazing. I like it. But you might say, okay, we have skills, but we also have hundreds and thousands of skills. We have skill libraries. And we still have a problem here because we
Speaker A
if we need new skills, we go to the skill libraries and the skill library become now a memory management problem for our agentic system. No, because skills cannot accumulate forever. They cannot eat up our token. So the new
Speaker A
memory that we have dynamically to construct, Google tells us in this paper contains now of the existing libraries that we need minus the skills selected for removal plus the validated new or the revised skill structures. No.
Speaker A
Now the external memory is great that we have it now. It avoids you overriding the m weights and catastrophic forgetting and whatsoever. But remember the external memory introduces a new class of failure modes. And this integrates here or this includes here
Speaker A
the retrieval pollution, the duplicated skills, the contradictory processes, the stale APIs, the dependency conflicts, the growing token and latency cost that we incur and all the activation collisions I showed you in one of my last videos. So therefore what is the
Speaker A
solution? The external procedural memory structure must therefore be created, managed, optimized and taken care of.
Speaker A
So Google tells us now in this paper, okay, a skill must earn its place in the final memory. So let's do this. They have here a simple mathematical function that they can code with. This is here the decision parameter. And they say to
Speaker A
determine which skills should be retained or evicted the frameworks evaluate now an empirical utility function. This is the classical idea. If we have no idea what to do we go with a utility function. The utility function in our case is simply here. The first
Speaker A
time the first term is here you see a counterfactual performance gain of injected skill s over minus the base policy pi base.
Speaker A
And then we have of course another term and these are the costs. So the costs plural that incorporates the retrieval latency, the token footprint and the execution overhead and everything that is associated with those costs. And they say if we are now above or below a
Speaker A
certain threshold, this will decide here on this particular which skills should be retained or evicted.
Speaker A
And they tell us, you know what? Now we have a new condition which skill to select because the skill should remain in our narrow task library only if its marginal benefit is positive.
Speaker A
And you say okay this makes sense. No, because remember a skill can be a valid skill but produce such problems on different fronts that we don't want to have it. So they tell us skills whose empirical utility that we just
Speaker A
calculated falls below an eviction threshold or whose activation frequency exhibit here a temporal decay function are now systematically purged here to maintain a library solveny. our beautiful powerful library and as I told you skill can be individually absolutely correct and
Speaker A
still reduce the global agent performance or the agentic performance of your multi- aent system because it may be retrieved too often. It may have a conflict with a better skill add excessive context invoke expensive tools cause the router to confuse neighboring
Speaker A
procedures and whatsoever. So therefore the unit of evaluation of skills and which skills from the skill banks or skill libraries it must be evaluated on its behavioral contribution and not just of its textual quality and this is quite a move quite a new
Speaker A
paradigm here for skill evaluation for your particular task. The next step is that Google tells us in this presentation, you know what's the stored competence here in the skill libraries is not an accessible competence.
Speaker A
And you might smile and say, yeah, this is a classical rack topic. No, so we have here our top K from the semantic similarity and our BM25. This is the classical boring stuff. And then we have a reranker like rack and yeah, you know
Speaker A
this. And then we say Google tells us here and therefore the effective competence here is something this is interesting and have a look this is now not not an addition this is here multiplication and Google tells us you
Speaker A
know what it is here the competence here is the stored skill quality times the retrieval quality rack times the execution reliability oh wow this is a heavy parameter times the verification quality and another heavy parameter So suddenly we don't use skills like hey
Speaker A
yeah let's do this skill the skill has in the title something that sounds semantically similar I have a cosign similarity in a in a vector space so let's go with this skill oh no way we have the execution reliability in the
Speaker A
complete system integration and we have the verification quality as a test object here in the complete system if any of [clears throat] these factors remember multiplication approaches here zero goes to zero the complete capability fails. So this is now a nice
Speaker A
quality benchmark by Google for the effective competency of the skills. And finally, finally, if you go over all the pages here, you have the insight. Okay, this means a skill library needs an operating system. You just can not have
Speaker A
here just piles of skills, an operating system. And you might say, but wait a minute, why? Well, once the skill library grows now, the harness must now provide what? The complete discovery process of the of a particular skill in
Speaker A
the library, the complete ordering, the complete representation, the complete storage complexity, the complete indexing and not only the semantic close indexing, the complete retrieval process and the complete routing process. And you might say, "Yeah, but I can do this
Speaker A
here with multiple agents. I don't need an operating system." And I say, "Yeah, beautiful." But this is was just the beginning because we also have to take care about here the harness has to take care about the dependency management the
Speaker A
composition complexity the execution complexity the complete verification system the repair system I mean hello baby the complete adaptation the evaluation which goes to infinity everything it's new about security and versioning and decommissioning on everything so okay you might say okay
Speaker A
maybe we really do need an operating system and we do follow here ghoul's advice for the harness.
Speaker A
So this means now converting the aentic skills from what it was up until now an isolated prompting technique. No, a skill markdown file, a memory markdown file, a little bit of examples, a little bit of something else. Yeah, beautiful.
Speaker A
Well, that stops now because now we have to integrate agendic skill into a full system life cycle because remember skill is both a knowledge a knowledge complexity and an executable supply chain.
Speaker A
So skills I mean beautiful on the one hand side know they can call tools they can execute the code they have access credentials they can read all their files in your system in the memory in the cloud whatever they can modify all
Speaker A
the environments they can invoke other skills they can if they by open eye here attack here hugging face server and have your absolute unbelievable dynamics so therefore they also become suddenly attack surfaces because yeah they are out in the And therefore there's some
Speaker A
malicious code and they're just waiting for our little agents with our little skills come out to play and Google lists us here. Thank you Google here all the security risks. So we have here external modified viability. My goodness that enables
Speaker A
tampering. Dynamic retrieval enables your malicious routing. Composition enables privilege amplification. Natural language instruction enable prompt injection. Executable assets enable direct system compromise and MOT places great supply chain risk. Oh great and [laughter] we could cry because yeah absolutely unresolved.
Speaker A
Google tells us the defense would require here complete provenence signature static inspection sandboxing runtime monitoring post condition verification network restriction lease priv credentials and behavioral evaluation at all moments in time. So you see the skill is now the point where
Speaker A
just a what used to be semantic extraction become now really executable authority. My goodness you know what started here some days ago so beautiful with some wiki skill by Google as a skill compiler. But think about this now if
Speaker A
you read the paper this paper here this preprint by Google you will see that wiki skill it really glides into this system. This is in no contradiction at all with this new idea in this new preprint notation. The wiki
Speaker A
ski wiki supports here the discovery operator we talked about. It supports the skill proposer that produces now our candidate s the validation gate. Remember if you have not seen my video here on vicki skills we have there the validation gate
Speaker A
implementation and the accepted result of course enter here particular memory structure here the developer memory.
Speaker A
Yeah. So this means this new system foundation preprint that we talk about today provides here a beautiful formal location for the positioning of wiki skill. Wiki skill is not alone anymore in the dark and naked it has found its
Speaker A
place. And you know what the other video where I talked about the dual memory RSI the self-improvement the recursive self-improvement here this also finds a place. Remember we had a dual memory paper separates here the experimental memory the reusable skills
Speaker A
and the procedural lessons from the working memory. These are the current unresolved goals and the verified execution state. Guess what? In this new framework paper that we just talked about the experimental memory corresponds largely to the procedural library. The working memory contributes
Speaker A
here to the current state. The invocation policy that we discussed in the paper here on RSI implements part here of our activation and the retrieval mechanisms. The checker evaluate whether the intended effect actually occurs and of course the meter agent modifies the
Speaker A
skill on the memory components after failure and we are again compatible with this system.
Speaker A
And you may say okay two videos fit but what about one of the latest video on skill state?
Speaker A
Well, guess what? Who was the author of this paper? Huh? Skill state addresses what happens here to X during a long procedural episode. No. And I said instead of repeatedly appending here the complete controvers conversational history, skill state by Google showed
Speaker A
that it can maintain here an explicit structured state condition. And guess what? Skill states supplies here concrete implementation of this state transition substrate. And today we are talking about the state transition operator.
Speaker A
And you might say what a coincidence if you have not looked at the mathematics of the paper. So remember skills state ask hey what exactly is x of t and how can it be updated without accumulated here again the complete transcript
Speaker A
and therefore we can deduct and yes I have a little smile on my face and say through a conceptual lens and only if you're a subscriber of my channel you will understand this hint of the conceptual lens you see also who or why
Speaker A
this systems foundation paper defines now the skill abstraction so the skill state defines contains a scalable runtime state architecture for executing it. So you see skill state well Google is compatible with Google.
Speaker A
Now you know if you are one of my channel members here the expert group here on my channel you know this is a video here from days ago where I showed you skill does not equal to authority and I showed you the new auto policy for
Speaker A
skills and we can even integrate this video because the auto policy belongs guess what to authorization on actuation boundary remember auto policy develops here the question which action is actually allowed under what lease and I explained what the lease is in the video
Speaker A
or authorization is a particular action by the LLM allowed. At what moment may the action cross into the external environment and how can malicious or misalign instruction be prevented from reaching here the actuation and this was the main topic here of the video for my
Speaker A
members. Yeah, you remember the connection was that the skill relevance does not imply an action authority and yeah there was even the paper I think when you see this of yesterday of wise openg rails and this was about how runtime controls
Speaker A
compose around the execution. Now remember I showed you in the video what are rails here and this is not Ruby on rails or anything. This is rails belong here to the orchestration and the runtime governance layer. So they
Speaker A
intervene around the model calls the tool calls the context construction the stopping the reflection and the error handling as shown in my last video.
Speaker A
This means remember a skill tells you an agent what procedure should be performed and reals here from my last video help determine how that execution is mediated, observed, modified or even stopped by the harness itself in the harness development phase.
Speaker A
Now having now all these last four videos and having now this particular video we can dare giving here this new paper by Google to build here a four layer unifying system approach but this is really the leading and bleeding edge
Speaker A
of AI because this was published today and I produced this video on today September 1st. So yeah, there might be some logical inconsequences, but I'm just enjoying playing around with these ideas.
Speaker A
You can say the Atlantic skill stack. We have four elements. We have four layers.
Speaker A
I mean, yeah, the inner layer core LLM is not a layer, but you get my idea.
Speaker A
So we have here our neural network at the center, the core LLM interpretation of the complete planning cycle. Then we have the runtime state where we have our goals, our progress. Then we have everything that we talked in the last
Speaker A
four videos about skill memory. Remember this is now one term about retrieval execution. And then we need everything about authorization, verification, cyber security, governance and execution harness levels. Beautiful. And now I say let's build something that has just four
Speaker A
layers. I know there are much more complication behind this but let's go with four layers.
Speaker A
So remember the LLM responsibilities here. Never forget the LLM because this is the main brain here of the agent is responsible for the complete interpretation of the my query complexity is responsible for the complete AI reasoning given my query
Speaker A
given here the information provided by the context window and given here the data provided by the harness complexity.
Speaker A
The LLM is responsible coming up with a plan how to come closer here to how to analyze the problem, how to maybe find a solution to the problem. The LLM is responsible for the semantic generalization and for the production of candidate
Speaker A
actions. So huge set of responsibilities. But let's look now at the other layers, the other three layers if you want. Now you immediately see that the runtime state as defined in this paper by Google integrates our skill state here from two
Speaker A
days ago. the skill memory let's say layer here integrates here Vicki skill and the dual memory evolution in a beautiful way and the governance and execution harness now itself integrates now the auto policy I know it was a memo
Speaker A
video and the new open go rails by UI technologies so you see everything falls in place everything becomes now more or less an interactive interoperable system dynamics.
Speaker A
Yeah, at the end of my video I always have imitations but let's say what is nice on this paper and I like this paper and there are some scientific strengths.
Speaker A
No, they have now a new definition a new boundary across here their term skills and they have this unified six component abstraction here as a six component tuple. Okay, if you go with this, we can fit here the other papers that I showed
Speaker A
you last week beautifully into this framework. So I kind of like this idea by Google.
Speaker A
We have built now Google has built a life cycle oriented complete system model a connection between the procedural memory that we had last year last week and the execution level the explicit attention to the routing the verification and the security aspects
Speaker A
and it is kind of a synthesis of a fragmented field bringing everything together that nobody else has brought together yet. So I like this approach.
Speaker A
However, however, there are limitations and let's be clear, there's no unified implementation of the proposed architecture. There's no control experimental testing here of the complete life cycle itself. There was no causal ablation establishing which stages are necessary and how they
Speaker A
contribute here to the result. There's no proof that this six double is really sufficient. Maybe there are other parameters or there's a different complexity level we have not addressed yet in this formula. There are no universal semantics across the natural
Speaker A
language executusible and graphbased skills. There's no complete solution to the routing and no proof that the that the declared permissions match actual runtime behavior. And yeah, maybe I have missed some points but you get it. So you see this is another step forward,
Speaker A
another leading and bleeding edge of AI research of EI technology development that is absolute fascinating.
Speaker A
But what is really nice and what I wanted to show you in this paper, you can integrate here the latest development into the system approach.
Speaker A
Again, wiki skill explains how the experience can be compiled into a skill. Have a [clears throat] look at this video if you're interested in this topic. The dual memory video explains how the current state selects and improves a particular skill set.
Speaker A
skill state explains how the execution can proceed without accumulating an endless transcript and filling up the context windows of your LMS and auto policy. Finally, the constraints which actions may reach here the world where we have here really the action execution
Speaker A
and the feedback by the environment based on this particular action set. And this new preprint of today by Google reveals that these are not just separated tricks or elements or fragments. No, these are or these are becoming now components. If you read the
Speaker A
paper and please have a look, please do it. They are becoming components of an emerging system layer around the core LLM of our agent. So the harness is in a massive development. The harness components are developing. They are
Speaker A
merging together but they're at the same time entering a higher complexity level. We have now as I shown you with the utility function and the utility cost structures here complete new thresholds where we say we don't accept any skills.
Speaker A
We have to have the verification within the system boundary within here a validated test bed with some particular tools a set of tools that are verified that are really optimized for this particular case and you got it. So this
Speaker A
component development of an emerging system layer wow the development is really breathtaking. So keep up the pace because read as much as you can inform yourself what is going on and I think this paper is a beautiful place to
Speaker A
start. So therefore at the end now finally we can say okay with our beautiful new knowledge we know the LLM remains the semantic engine logic decision action beautiful but the skill system now around this LLM determines which procedures survive when they
Speaker A
skills appear what they may control and whether their effects are accepted by the harness control cycle. So this means the future agent is therefore not simply another smarter AI model because it will be a model that is deeply embedded inside an evolving
Speaker A
procedural operating system and not anymore a procedural memory string. It will be a procedural operating system integrating everything that we talked about in this video. And this development is absolutely beautiful and I guess it will lead us here in an
Speaker A
absolute fascinating new direction of AI technology. I hope you enjoyed this video, had a little bit of fun, maybe some new ideas, maybe something you want to try out. Anyway, I hope to see you in my next video.
Topics:agentic skillsAI procedural operating systemGoogle AI 2026state transition operatorskill lifecycledynamic skill invocationskill modularitymulti-agent systemsskill memory managementAI system architecture
