Speaker A
Hello community. So great to see you. Yes, you're not going to believe it. We have a new object in artificial intelligence. If you are really want to be at the leading and bleeding edge of AI, welcome to proidual graph
Discover Google's new procedural graphs for AI, enhancing control and decision-making beyond traditional knowledge graphs in LLM agents.
Key Takeaways
- Procedural knowledge is distinct from factual knowledge and critical for reliable AI action over time.
- LLMs alone struggle with procedural tasks due to statelessness and lack of explicit procedural memory.
- Procedural graphs provide a structured, external control layer that guides AI agents effectively.
- This approach improves AI performance in complex, long-term decision tasks without modifying LLM weights.
- Graph-based procedural representation is essential for overcoming limitations of chronological memory in AI.
What the video covers
- Google, Georgia Tech, and Peking University introduced procedural graphs as a new AI control layer.
- Procedural graphs organize procedural knowledge into triplets, unlike knowledge graphs that store factual data.
- Current AI models, even with vast knowledge, fail in long-term procedural tasks like financial management due to lack of procedural memory.
- Procedural graphs provide a structured, editable map guiding AI agents through complex tasks over time.
- The approach keeps the LLM weights frozen, adding a procedural graph harness for improved reasoning and decision-making.
- This method addresses failures in long-horizon tasks by explicitly representing procedural sequences rather than relying on chronological history.
- Procedural graphs include conditions, guidance, and pitfalls on edges to localize decision-making.
- The graph structure allows the AI to focus on relevant local context rather than entire task history, improving efficiency.
- Google's paper demonstrates improved performance on benchmarks like enterprise financial management using procedural graphs.
- The procedural graph evolves through self-repair and expert priors, enabling continuous improvement.
Chapters
- 00:00Introduction to Procedural Graphs in AI
- 02:19Procedural Graphs vs Knowledge Graphs
- 04:49Limitations of Current AI in Financial Tasks
- 07:15Procedural Graph Structure and Components
- 09:21Graph Localization and Guidance for AI Agents
- 13:57Comparison to Skill Markdown Files and Procedural Representation
- 20:26Mathematical Formalism and Harness Integration
- 29:35Graph Evolution and Performance Improvements
- 35:44Summary and Future Directions
Full Transcript — Download SRT & Markdown
Speaker A
structures. And you might say, who dares to come up with another idea? Well, guess what it is? Google, Georgia Institute of Technology and the Peing University. Now Google and Peing University. This is here constellation.
Speaker A
I do have to show you this paper published September 8, 2026. Protoidol graphs a new paper by Google about self evolving execution structure for LLM agents.
Speaker A
And they have a simple example why we need more of AI. Look, there's now the intelligent CFO, the financial officer of a corporation who still goes bankrupt every time. So we have an enterprise arena. This is here benchmark failure.
Speaker A
So the eye agent knows finance knows every has read every book on corporate finance. It sees actively here the agent that the cash is falling. It understand that the bankruptcy is coming yet almost all the agent wait until the month 32 in
Speaker A
this particular benchmark to request additional funding even though the funding takes here one to six months to arrive. And now Google said the key question is why why the hell is it possible that an intelligent AI model understands the danger has all the
Speaker A
reasoning capacity in the world runs on I don't know 100 GPUs and is still not able to understand that it is acting too late now it turns out the answer is intelligence even artificial intelligence is not procedural knowledge
Speaker A
so knowing facts is absolutely different from maintaining a reliable action structure within our agents.
Speaker A
So therefore says hey let's introduce the protool graph. Now remember what is a knowledge graph. The knowledge graph organizes factual knowledge in triplets entity relation and another entity head tail relation as I showed you in my last video for this what is questions. And
Speaker A
now Google builds a procedural graph and this organizes procedural knowledge into a triplet structure. procedure one the relation and procedure two for what to do questions and immediately you think about wait a minute where we have skilled markdown files now kind of a
Speaker A
procedural knowledge so what is happening now I will show you that across task and model family this new methodology delivers the absolute gains over the memory baselines plus at the end we will talk about of course now we
Speaker A
are looping so we will do have we have to have a self evolution structure of our procedural graphs. We will build effective graphs from a minimal initialization and repair expert prior that initially hinder our performance.
Speaker A
So here you have now the official um drafting. I take a screenshot here from the paper by Google. Now here the knowledge graph where is the monola and you have here the protoidal graph. What should the agent do next? give them a
Speaker A
complex problem. Now it is really important in our understanding and if you are a subscriber of my channel you know exactly why I stress this the LLM remains frozen. We do not modify the learning of the LLM itself. We do not
Speaker A
touch the tensor weight in the transformer layers of the LLM. What learns is here a component of the harness and this component in the harness is the protodual graph surrounding here the pure LLM our reasoning engine.
Speaker A
Now again think about this a highly intelligent CFO AI agent who understands finance everything but it suffers somehow from procedural amnesia. It knows everything but it does not manage to come up with enough resources. So every month the CFO receives the
Speaker A
complete history of the company and can reason intelligently and you know exactly what it means. We have a context window and because LLMs are stateless things we have to provide the complete history of a complex task everything and
Speaker A
then we will have one reasoning step in a next token prediction um LLM yet the CFO still checks the same balance repeatedly waits too long before raising additional capital forgets that the funding takes several months to arrive. The company goes bankrupt.
Speaker A
submits a second request while the first remains pending and performs one additional action after the task is already complete. So this is the current performance of the best AI models we have in finance. And you might say this
Speaker A
is not okay and Google tells us but we have a solution. So okay let's have a look. I mean this insight I mean even my human learning what I do as a human I sometimes learn what we implement now
Speaker A
new algorithms for the machine as I told you the problem is not missing factual knowledge the problem is maintaining a reliable procedure over time and this is a simulated 132 month company so a conventional agent receives here the task description
Speaker A
and then as I told you a long trajectory what has happened until now what were the action what were the environmental states and then you have this as an input to the L&M and the LLM now predicts the next action this is our
Speaker A
basic pattern how an LLM works with an harness or without an hornice so the LM must reconstruct the procedure from the entire history at every single reasoning step at every single context window and now Google says stop this is just
Speaker A
nonsense we We have now a procedural graph. This is another graph structure. This is not a knowledge graph and it's the stores here something like okay we have the month start we check for the cash we have a forecast runaway we check the
Speaker A
market condition and then we decide if we need additional financing. So you see every transaction additionally contains here three elements. a condition when this transition applies, a guidance how to execute it and the pitfalls what errors to avoid and I would say great.
Speaker A
So the agent is now localized somewhere on its current node shown only the nearby graph structure a two three hop element and give my situation specific guidance or its next decision. So the sortest description here is something like this. A procedural graph is an
Speaker A
external editable map of how an agent should move through a particular task should move how to manage the cash flow of a company here for 100 months. So you see this is closer to a learned operational doctrine than to an ordinary
Speaker A
memory structure. I think this is the power of this paper. Think about it. A classical react style agent usually conditions its next action on the accumulated history. Now this is here the history. This is our trajectory. We have here an action one, an observation
Speaker A
one, an action two, an observation two, two t minus one steps. Great. But this history that we have now tells the AI model what happened in the past. But it does not explicitly tell the model where it is currently inside the larger
Speaker A
procedure, which prerequisites have been satisfied, which steps remain for the graph completion and when the task should terminate.
Speaker A
So there is a significant difference because the classical trajectory that we operate until now is chronological but it is not procedural.
Speaker A
So even a long transcript where we have all the information in it for the context window of the LLM might contain all the necessary information yet the AI model must repeatedly reconstruct the hidden procedure from that transcript every single time and this is just
Speaker A
boring even for a machine. So if you wanted papers deeper hypothesis is many long horizon agent failures are failures of procedural representation of a particular sequence of action not a failure in the raw intelligence of our LLMs. It is not fable 5.1 or whatever it
Speaker A
is we are missing something and pros dual representation and Google goes with a graph representation.
Speaker A
This particular distinction is here scientifically important because more context does not necessarily solve the problem at all because the context remain an unstructured log event.
Speaker A
Therefore, Google goes for a graph structure. Why ordinary memory is insufficient? Why do we have to go with a procedure world graph structure? Think about it. This is my sort. An episodic memory system may retrieve the previous run requesting funding early was really
Speaker A
helpful for the cash flow of the company. But the model must infer still infer where that lesson fits into the current sequence of the complete 120 month enterprise operation. No. And now a procedural graph instead represents here the data simple. We have a forecast
Speaker A
runaway and this leads to in the triple to a financing decision. It additionally tells us again this pairs of three condition guidance pitfall the condition the projected runway is below the safety buffer guidance request funding early because the delivery is delayed 1 to 6 month and
Speaker A
the pitfall do not submit another request while one is already pending. So the graph does not merely remember an experience like a skill or whatever but it places here the lesson inside an action topology.
Speaker A
So we do have a new mathematical complexity a topology that is here an action topology and we provide a guidance here by the graph structure and as you seen in my last video we do not just go for here a simple note and note
Speaker A
to node or a link to link relation but we will integrate the context we will integrate an subgraph topology for additional context and yes this is kind of a continuation to my last video and demon most important or the most complex
Speaker A
AI paper in 2026. Pearl on the other hand I ask myself but wait a minute please read the paper it's absolutely fast a lot of details but I ask what exactly is new and you know what none none of the individual
Speaker A
ingredients none of the things that we work with is entirely new agents already use textual memories workflows already encode ordered steps tool graphs already connect the tool calls conditional guidelines already say hey in this state do ABC an optimization system already
Speaker A
modify the prompt of the workflow. So all the elements that we have to work with or combine with or you get it yeah are there there's nothing that we have to invent but the choke is the protoedural graph itself is a procedural
Speaker A
invas in innovation. So the novelty is the combination of four properties. I know it's dense but I want to show you all four here in one glance. Explicit procedural topology steps in admissible transitions are represented as a directed graph. Attributed transitions
Speaker A
the conditions guidance in pitfall R3 live on the edges localized generative guidance. This means the AI model receives the neighborhood around its current procedural position rather than the entire graph because this would be the other option. I give you the entire
Speaker A
graph of millions of object. No, I overflow here the context window here with an immense intensity of data. This is absolute nonsense. We have to have a selected subgraph and the feedbackdriven graph evolution. Successful and failed trajectory modify the topology and
Speaker A
attributes subject to a held out validation gate. I will show you the formula for the validation gate in 5 minutes.
Speaker A
So here we go. A knowledge graph. Oh, beautiful represents factual relation. Mona was painted by Dainci great triplet.
Speaker A
Acknowledge graph answers what kind of problem. It answers what is true. But what the folin does here? Our procedural graph answer. The procedural graph answers what should happen next.
Speaker A
It represents here an action relation triple. Let's say search leads to extract particular evidence.
Speaker A
This is the way I see it. Now maybe I'm wrong. Please correct me. So whatever you do not see here on the left side here the indicator of the document you know those are my ideas and this is how
Speaker A
I try to explain a certain complexity to my viewers. So knowledge graph the node represent the entities the edges represent here the factual relation and the primary question a knowledge graph wants to answer is what is true proal
Speaker A
graph and now now you're going to be amazed I think the note represents the action the skills or the complete state of a task the edges represent the admissible transitions the primary question here for a protoal graph is
Speaker A
what should happen next and now We come into a topic that I ask myself. Wait a minute. What is the difference to SQL MD file? But a protocol graph and note if it can represent an action or a skill
Speaker A
description or a complete state task description. Where are we? And I think a protoal graph is a representation of a protoal memory complexity rather than a semantic memory representation.
Speaker A
Why? So let's talk about procedural memory versus procedural growth. A procedural memory and I have to be really clear in my definition within my videos because I want to be coherent in all my videos. So procedural memory means knowledge about how to act, what
Speaker A
to do, in what order, under which conditions, how to avoid certain things and when to stop.
Speaker A
So procedural memory can be stored in many forms. natural language instruction or a skill markdown file or if then rules or state machines or workflow or procedural graphs or an complete higher step learned neural policies itself.
Speaker A
So you see careful procedural memories and procedural graph what is it to how can it be stored.
Speaker A
So this brings me now in the definition and the quest for definition. What is now the definition the mathematical precise definition of a procedure.
Speaker A
I think a procedure is the actual strategy or the operational pattern for a particular task. A procedure could be eight steps. Formulate the research question. We are here in science.
Speaker A
Generate the search terms. Search for papers on the internet wherever. Evaluate the source quality. Only go with peer review or also integrate preprints. Extract the evidence, crossch check important claims, synthesize the results, and stop when the evidence is
Speaker A
sufficient. You have a certain threshold parameter. Eight step one procedure. Okay. So, let's go with this procedure. An operational method. The main question, how should the task be performed? Skill a packaged agent capability. The main question, what capability can the agent invoke? Skill
Speaker A
markdown file a textual file representation of a particular skill. How is the capability described and loaded and procedural graph now is as you see skill to skill MD and procedure to procedure graph as pro a procedural graph is now an explicit topology of
Speaker A
actions or procedures. The main question here is where is the agent here on our not knowledge graph but protodual graph and what should happen next if I follow here the graph topology.
Speaker A
Now I think and now it gets crazy I know but stick with me a graph can organize now either one skill or many skills because I even can have a graph inside one skill. I operate with this. Huh?
Speaker A
Literature review skill could contain an internal procedural graph connecting its individual actions. Search, evaluate, extract, verify. I just showed you an eight step.
Speaker A
But a graph can also connect multiple skills into a super skill. If you want at a higher level, every node might now represent an entire skill itself like literature search skill to a PDF an analysis skill to a citation
Speaker A
verification skill to a scientific writing skill. So careful with the terminology. I think a procedural graph is a graph structured representation of procedural memory. A skill MD file is a package textual description of a skill that may itself contain procedural memory.
Speaker A
And I think the preprint advantage that it is shown is not that it invents now procedural memory beyond skills. what I thought about when I read it here the first time. I think what it does, it really converts procedural knowledge
Speaker A
from a largely textual instruction markdown into an explicit localizable and independently editable control structure of a graph representation.
Speaker A
And then I ask myself, yes, I have a question a lot of question when I read a paper. Why the preprint does not simply use skill markdown file? Think about it.
Speaker A
A static skill markdown file usually provides only a textual instruction, but the LLM still has a job to do. The LLM still has to infer in addition which step it is currently executing on a graph. You see it immediately which
Speaker A
branch it applies, which instruction are locally relevant, you know, our subgraph topology, which steps has already been completed on a graph. you will know exactly where you are in this graph and when it should terminate and you notice
Speaker A
this exactly on the graph representation and I think the beauty of the preprint is by Google well guess what it externalizes this relationship it brings it into the harness and we have now a very intense LLM harness relationship
Speaker A
now if you say hey there's for sure some mathematics yeah but it is so simple look this is it whenever you see here this green thing here on the side. You know this is kind of the main summary of
Speaker A
the complete paper. So at a particular time t here this preprint identifies with this new methodology an active node on a protoal graph. This node is not v4 vertex but u okay this is here notation by Google we have to accept it. Then we
Speaker A
have a matching function. Then we have all the action up to t minus one. And here this is a particular set of vertices.
Speaker A
And from this it retrieves now a nearby graph at N2. N for neighborhood two for two hop uh graph theoretical neighborhood here of UT around UT. And then it generates guidance and this is the main part of the paper. We have
Speaker A
additional guidance from the local neighborhood. Remember in my last video on the most crazy paper in 2026 Pearl we had the same thing. We defined here a local neighborhood on a graph structure and then we had a particular one hop or
Speaker A
two hop or four hop environment that we redefined as the context for our particular task and we said hey is this node on this particular edge that we want to discover or not here now the formula is here five function and we
Speaker A
have a explain it in a second so now you understand here why I've chosen to bring up here particular image. Now we have a frozen L&M. We have here a protoural graph. We go for a local subgraph structure in a particular
Speaker A
neighborhood that we define. And then we will have a simple selection rejected or green lighted. Beautiful.
Speaker A
But think about it. The agent is not simply handed a large skill description. It is told, hey, you are here as an AI agent on this particular graph node.
Speaker A
These are the relevant next transition that are allowed for you for the execution of your particular job given and this condition currently applies and avoid here known failure here the red uh edges that you see that are around you.
Speaker A
So you see the kind of complexity for deciding what to do next is now encoded in kind of a deterministic graph structure and is not up to the hallucination of the LLM. Furthermore, the system can now modify individual
Speaker A
node edges conditions and pitfall based on experience and this will lead us to a self evolution of the proidual graph but this will be at the end of this video.
Speaker A
So let's go with the formula the matching. Yeah. A t minus one the previous action. V is a set of graph nodes. Great. This formula means simple.
Speaker A
At a decision time t compare the agent's previous actions a t minus one with the nodes in the procedural graph given its particular location in the protoal graph structure and identify the node that best represents that particular action.
Speaker A
You know graph protoal graph is a normal graph. So we have here a graph here. We have our vertices. We have our edges here. Great. And a note can now and this is now the beauty careful. A note can
Speaker A
represent the following objects. An action a tool call a procedural step itself a task state and sometimes as I showed you in my opinion an entire skill.
Speaker A
So let's have an example. A tus one is a a cash flow forecast calculation. And this might be matched to a forecast runaway prediction. So we have match cash flow forecast calculation a set of possible vertices and the result is
Speaker A
forecast runaway. UT I already told you is our active protocol node not a V for here uh vertices but here U and this is the note representing the agent estimated current position in the procedure itself. Now a lot of things can go wrong. This
Speaker A
localization can go wrong or the procedure graph is not really suited has not been optimized has not co-evolved for the complexity of the task. So you're operating in an incorrect mathematical representation of the solution space and two I already indicated it this
Speaker A
retrieve here the local graph it means retrieve the graph neighborhood of the procedural graph extending up to the two transitions to hop graph from the current node. Remember this is a pure mathematical graph network theory operation. This is a graph operation.
Speaker A
Beautiful. N2 can contain the current note for runaway one hop successor market two hop successor request funding transition conditions guide and pitfalls great and then we have our side and this is here simple formulation that tells us this is
Speaker A
our guidance generator and what is our guidance generator in the age of AI guess what it is an LLM call so the LLM call combined now with a prompt template depending on the complexity of your task. Here you can have here some
Speaker A
deterministic guidance too. It converts now a graph information and the trajectory context into the natural language advice. This is why we need the LLM here for our guidance generator. So you see we operate here an LLM on a
Speaker A
graph structure. Yeah. T, you know, represents the agent's interaction trajectory and T from T to T from T minus W to T is simply the recent trajectory segment.
Speaker A
You got it? So if you want to have it in very simple terms, the guidance generator combines now the following elements. A general procedure with the current goal from my task and the recent evidence that I have here in my
Speaker A
procedural graph. And this leads now to a specific advice situation specific graphbased node localization advice that is now if you want cooked up by the LLM.
Speaker A
So this combination of a graph position the local procedure the current objective and the recent evidence is here everything that is here the meaning of these three equations and then we have a solver. We have to have validation. We have to have a
Speaker A
solver and let the solver choose it. So the solver receives now the full task, the full trajectory and the generated guidance for particular topics and the complete online policy can therefore be this simple equation for our actions.
Speaker A
Great. Now I told you the whole time already there's a second part and the second part is if we have this graph this procedural graph now we understand what it is how to use it now there's a self
Speaker A
evolution part to this procedural graph and guess what it has to do with loops we have two different time scales we have a fast loop this is for the agent execution so within one task period here our graph of is fixed and while the
Speaker A
agent produces here action observation action observation using here now the local guidance by our now beautiful added protodural graph structure and then we have another time scale and let's call it slow looping this is now the evolution of the proural graph
Speaker A
structure itself so between our training batches the model weights remain fixed the LLM model weights remain fixed but the external procedure represent mentation changes.
Speaker A
If you want to see this in our mathematical harness notation here, the harness here operates now on the evolving or with the evolving procedural component G, our graph structure and the harness operates here on our LLM model M, our fable 5.1 or whatever you have.
Speaker A
Great. So you see that we have now a system that evolves here not from a model to a to fable five to a fable 5.2 5.1 and 5.2 two but the system evolves now here with this guidance the graph at zero graph at
Speaker A
one graph at two and you got it. So if you want this is what Google calls now an external prototural learning and this has nothing to do with the parametric model training of the LLM. This is happening now in the harness in our
Speaker A
network complexity. So self evolution for procedural graph. So we start an offline self-evolution loop adapts now the graph topology and the attributes from the execution feedback reducing here the need for any manual design or any human in the loops.
Speaker A
We have four elements. We have a diagnostic roll out, the feedback driven mutation. I think real beautiful the validation gate because everything has to be validated and then all the rejected memory. All the failed memory is not thrown away but stored here as
Speaker A
here in somewhere in a memory as kind of a safeguard to know what not to do in the future. But I think let's talk about the validation gate. The validation gate keeps here a graph modification. A graph it is not a genetic algorithm. It is not
Speaker A
an automatic evolution. It is just a graph modification. Keeps a graph modification only if it produces an equal or a better solver.
Speaker A
Not merely a graph that looks more plausible that is more beautiful or whatever. You have to have a criterium and this is equal or a better solver.
Speaker A
This is here the reason why we have this validation gate here at this point in time. So if you want a validation gate is a quality control checkpoint for changes to the protool graph itself. The refiner proposes a modified graph. The
Speaker A
solver uses here both graphs now on separate validation tasks. Their solver scores are compared which one is bigger and if the candidates perform at least as well it becomes the new graph otherwise it is rejected. As simple as a
Speaker A
gate can be. This is here the handwritten chaos by Google. This is a screenshot of the Google paper itself.
Speaker A
Um, yeah, I did not really get it here the first time. I had to read the complete text and then come up with this video results.
Speaker A
They ran a lot of tests and they ran a lot of financial company tests, a real simulation for AI agent in a real enterprise environment. I'm not really interested in this. I just give you the result of it. So what was the test? Four
Speaker A
different LLMs placed in charge of a simulated company for 132 months. Three economic crisis arrive at a particular time of 32,59 and 112 month.
Speaker A
But the agents are not warned in advance. So they just encounter a problem. Yeah. So the plot shows now in this simulation their cash trajectories and the upper plot will show you here particular survival curve representation where every downward step means that
Speaker A
another simulated company has gone bankrupt. The red line you will see is the ordinary solver and the green and the orange add some conventional memory that you know the blue line the blue line is now the beautiful new thing of our
Speaker A
procedural graph. Here you have it. What clothes on it? Survival rate 58% in blue. Cash blue is not the most not the best one. Look, the green is the best one, but the survival rate of green is just 50% and compared to a blue one,
Speaker A
blue is 58%. Gemini 3.1 Pro. You see here, I mean, I love it. They don't even use here a computer plot. No, they have here some handwritten something. So to make it absolutely not possible to reconstruct here this data. Now this is
Speaker A
handwritten graph structures by Google. I mean come on Google what are you afraid of?
Speaker A
We have here Gemini 3.1 Pro. This is such an old old model Google please. A Gemini 3.1 Pro. Horrible cash flow.
Speaker A
Horrible survival rate. 34% compared to 58% with clash. Yeah. just 23 million but the blue one survived. So okay Gemini 0.5 flash or Grock you got it.
Speaker A
If we look for the efficiency analysis and this is your table three and you have here the PG usage ablation with Gemini 3.5 flash. Let's go to Alford.
Speaker A
You see here the last line is here the new ours methodology and it almost outperforms everything else. Beautiful.
Speaker A
also the full graph and the full graph raw injection and the baseline without a graph it just outperforms. Great. But this tells us here especially if you have a deeper look at the details of AL world the strongest configuration is
Speaker A
that we have information about the local graph topology on a procedural level plus a situational generation.
Speaker A
If we have an understanding of the exact situation the agent is in and we know the exact place the agent is in on the graph topology and we have the local topology here for this particular position of the agent. The situational
Speaker A
information and the situational generation. Those are the most powerful information sources for a correct solution. You want to have the full numerical results. Enjoy it. And then you have here uh lifespan and capital raised through 10 rounds of self
Speaker A
evolution of here this graph structure. And so you have here for 10 generation here you see you have a beautiful okay in the first one in average survival month you have here yes some beautiful plateauing and average capital raised
Speaker A
here. Yeah, I'm not so much into this simulation here of an AI agent because I can think of quite a lot of improvement to this particular methodology by Google. But okay, I think really interesting are the insights. No, when should you use
Speaker A
procedural graph and this is now a new service here for you as a subscriber. I want to show you when to use it. What I think when to use it or not to use it. I think best suited here for repeated task
Speaker A
families for stable tool cataloges, multi-step dependencies, delayed consequences, important stopping rules, safety or policy constraints, recurring failure patterns, measurable validation outcomes, and sufficient execution data for refinement.
Speaker A
Good examples would be I think enterprise workflows, financial operations, definitely everything with scientific laboratory procedures, software debugging and testing, consumer service policy execution, think great in medical administrative workflows or autonomous tool use or everything that is connected with long horizon
Speaker A
simulation. On the other side, I would not use this system for if the task is simple one step. Creative freedom is more important than some procedural consistency. Maybe not. Or maybe if you're a bank office, you want to have
Speaker A
some consistency. If you're a creative office here or music, you want to have some creative freedom. Every task has a complete different action structure.
Speaker A
Then don't do not use this new methodology. If there's no reliable validation signal, if the tool names and interfaces change constantly. If you work constantly with different tool libraries, not recommended to have a fixed graph structure for this
Speaker A
interference cost must be minimal or a formal guarantee workflow is required. So this is it. But now take a step back and connect this preprint to the last videos and you will see what we have. We optimize our LLM intelligence as I've
Speaker A
shown you in last week. And now we add here this external procedural graph structure. And what we get out we get out a better system intelligence. So you can say this is an agent where we have the LLM at the core and the external
Speaker A
proidual structure is in the harness structure of this agent. So in total we have a better system intelligent.
Speaker A
But you know the protool graph is only one learnable organ here of the harness element of the harness.
Speaker A
stores operational knowledge outside the LLM model weights and can evolve much faster than the foundational model itself because the training of a harness is much cheaper, faster and I don't know better but okay what is the absolute beauty of this
Speaker A
paper I think the deepest insight is not that just graphs outperform textual memory structures but an agent procedural intelligence can be separated from its parametric intelligence and I will tell you more about application of this idea in one of my next videos.
Topics:procedural graphsartificial intelligenceLLM agentsGoogle AIknowledge graphprocedural knowledgeAI control layerlong-term AI tasksenterprise AIAI decision making
