Agentic Data Systems: Automated Knowledge Discovery

Key Takeaways

Hello community. So great that you are back. Today we talk a very specific type of agents data agents. And those are the research team members here. You see and you know it is all about data in the eye. No isn't it? And you know if you do data science 80% of it is some tedious work you have to do painstakingly labor here after data preparation and everything. And this new idea now is hey let's use AI about automating everything find the right data have the complete analytical reasoning process understand the right type of data visualize everything in the correct way so the main idea is data task are not simple operation but they require now logic especially a state management so data itself become now a state of a system and we add everything that we know from a chantic tool There's no complex tool integration. Every single one have robust, reliable AI agents. You know this. And if successful, the authors tell us those data agents could really democratize data science because allowing he a domain expert like a biologist or economist to perform their sophisticated analysis without really needing now some human expert coders who collects the data for them. So this domain expert if they have an idea they just go to a data AI agent say hey for this particular task find me on the internet on the SQL server no SQL whatever you have all the data of the world because I want to see if my new idea I can prove it with the data and the data agent will provide here the data for this so what is the product the product is here really clear it is a data science agent you can buy for example example for a specific domain knowledge like genomics. Now this data science agent has been particular pre-trained for the task has a supervised fine tuning and a reinforcement learning I'm going to show you in a minute or you buy a product a data science AI agent for financial risk assessment pre-trained only for this task and this agent knows exactly where to go on the internet what sources to find what data formats to look for convert what is available and what is the complexity here of the human task Because finding task specific data is not as easy as you might guess. Yeah. Because this agent has to understand the task, the human task, what the human wants and then it has to argue what kind of data that I know are existing on the internet or wherever should I collect, integrate, modify, bring together in SQL, NoSQL, whatever you have. And this agent has been done has been trained to do so is now really specific agent for the task. So you see it's not your standard I don't know data agent but this is now a data agent that allows everybody just to do your experiment and don't care where the data come from. You don't need a data generator here specific coders and the authors tell us this is for everybody to experience. So here we have it. This is the beautiful study September 23rd 2025 autonomous data agents they say hey there's a bottleneck we need data for AI you don't have to care if you are a domain specialist genomics risk management financial whatever medical we have the agents for you and then you have the data and then you do your experiments on those data new opportunities for smarter data and they develop your kind of what they call a formal framework data as a dynamic state itself. So what they say we have an action generation an agent responsible here who represented by LLM base function observes the current state of the data the schema the sample rows the meter data and its short-term memory has has the log of all the previous action and then it generates simply the next action and you say hey I'm familiar with this of course and if I tell you the next step is a state transition you will say hey I know this of course you know this so therefore the generated action A is now executed in a specific data environment, Python runtime, Panda, SQL database, whatever you have, which upstate now the data state from S1 to S2 or the next time interval and you say wait a minute let's have a look at the flowchart because if this is it what I have you have here an Excel sheet and then you just do manually if you don't have a data agent you just have your SQL and you have your select group by and order by and then you have to process on the visualization you use pandel metplot lip or whatever and then you have it and you might say but wait a minute I can code this no I can go with I don't know GPD5 or whatever you like yes of course you can code it but here you have data agents that have been particularly trained here for particular domain specific knowledge and they tell us here you know at first perception planning decomposition and then 1 2 3 four five six and they say those six tasks those subtask no we trained our agents for and here we have it again 1 2 3 4 5 six you see identify the data score generate a SQL command execute the SQL command verify the results here with I don't know GD5 do the visualization do the summary and this is it and the beauty is you could buy those agents know those data agents you don't have to care about how you generate the data the data are somewhere available mobile just do your experiment and the data are provided to you like data database SQL NoSQL API web services file CSV JSON why should you care as a scientist about the file conversion let AI do this no the data task that you have feature engineering symbolic equation extraction text to SQL automatic data repairs leave this to the eye you have now autonomous data agents here in their core what they do well it's very easy. They have a perception. They understand the data. At first they have to understand the task and then they look for the what kind of data they need. Then they do a planning and a decomposition of the complexity and action reasoning. They decide here on the action sequence. At first we have to clean the data. Then we have to concatenate the data. Then we have to filter the data. Then we have to do the grounding of course calling APIs here for validation and then the execution. And then you have here of course here the refinement the feedback and you know we're in a loop here of our data agents. Now, of course, you might say as a subscriber of my channel especially, hey, this sounds so familiar. Yes, I know this is always the same in AI. No, but imagine now. Now, if you have a real world data set, can be gigabytes or terabytes or even huge data set somewhere. No, and the agent cannot fit here the entire data set in the context window. So, it just has here a partial observation of the true state and said, "Oh, no, I know what's coming." Yes, of course here. So what the agent actually sees is a database schema, the first five rows here of a table with some statistical summaries, column names, data quality reports. And you know that the agent must now use these limited observations scanning all the data sources on the internet along with the memory of past action what it learned on the observation the belief state and now we say okay now this is an exact definition of a partially observable mark of decision process and we are back to what we know from AI. This is the cord that runs inside every large language model vision language model. This is it. So we are back to basics. Agent must act under uncertainty based on incomplete information about the SQL database making it a problem significantly harder than a fully observerishable market decision process. Now you have hidden states and you know everything about it. And at this point it said hey wait a minute wait a minute. So I read this study and I say but listen if it's just the a partially observable mock of decision process what is new about this paper it's ever this is just a standard application maybe some of you might say hey I just can use cursor and use this no and code this well it lies in the complexity I discovered no it's the absolute staggering scale of the problem space that you have to deal with if you're partial observable mock physition process which pushes here all the traditional reinforcement methods here to their absolute computational limit because think about it you have an infinite state space the number of all possible data set in the states that they're in is almost infinite I think at least computational infinite and then you have an action space that is a high combinatoral action space it's not just like up and down or whatever no structured compositional astronomically large space for syntactically correct actions and of course think about the reward function that you need a complex reward function because you have I don't know let's say 20 action so which one was the crucial reward coming back the complexity just goes crazy here if you do this but in principle if you see this from a theoretical point of view you say hey wait a Right. So it is the same basic idea. You have an agent and the agent goal is to learn here policy pi data and be fine tune here the llm or the agent here that generates now a sequence of better action here to maximize now the cumulative future reward. This is by definition still our old friend the reinforcement learning problem. Of course we have nothing else in AI currently. It's always the same. But of course to write a paper no we have now five pillars here of a data agent as defined here by our orers of this beautiful study. First is I make it real simple perception seeing the data. So you have data as tokens serialized tables or data as a visualization where you have to interpret charts or have to read plots or you just have a unified JSON structure that you have then you have the thinking process before you act at all. Yeah. So what you do with thinking you know LLMs are not good at thinking. So what you do you chop it up. You say I reduce the complexity to a lot of very low complexity subtask. So we just break it down into a logical plan of smaller executable subtask like data cleaning and then feature engineering and then aggregation then grouping and then visualization and forecasting. We break everything down. Great. Then we have to choose the right tool. No. Oh, and we have tool calling whatever you like, pandas or SQL or whatever you go, symbolic expression and a direct generation of insight on everything. And of course, now grounding action, grounding and execution. So you really have to have code, executable code that works here with the actual data scheme with the data flow that comes in and all the API signatures compatible beautiful coherent structure and you just have a learning from the feedback the refinement great if you have a SQL syntax error you know you have your self-correcting system if the intelligence of your coding LLM is huge enough they tell us you know what we our task are data processing feature engineering data augmentation, data visualization, text to SQL, data to equation and tool calling and all the things you have here. Beautiful. And they say we train our agent on those. We want that our data process that our data agents are really smart and autonomous. So it is the training process that you say okay so this is not just that I sit down and I code with course and I code sequence and then I have an agent. No, they say listen, our agent is now specifically trained for an extreme complexity of data task, data science task. And how we do this, you know, in the eye, it's always the same. We start with a supervised fine tuning and then we go to a reinforcement learning. And guess what? We start with a supervised fine tuning and then we go with a reinforcement learning also in this paper. So teaches you the agent the basic skill of task decomposition, skill invocation and tool use in a supervised manner. So you provide the supervised fine-tuning data sets and you can do this. Now of course the complexity would be too high if you go with too many agent and here in this paper you remember it's just a university they introduce only two architectural variants a single agent and a dual agent. And you may say oh wow yes I know it's oh wow. So single agent guess what single monolytic LLM learns to perform all function from planning to execution or you have planner agent and an actor agent and you have the feedback loop so basic stuff nothing complex it's just the amount here in the training data now comes the beautiful part reinforcement learning and they make here the simplest approach that's possible now for the policy optimization to go with PO the proximal policy optimization code that is available I don't know in 1000 variants so no problem at all you just need the data the data sets for the training and this is the beauty of this study and of course you have to have a reward function so you need a reward model and guess what you can even go with binary reward structure and you have to care about the execution success the correctness of the solution and the logical coherence in a complete execution here of this linear plan. That's it. Here you see it. Single actor and here you have it with the planer. That's it. There's nothing fancy. There's nothing special. It is just applied here to have now a pre-trained, supervised, fine-tuned and reinforced learned data agent that is highly let's say clever not to say intelligent for a particular domain knowledge like genomics, chemistry, theoretical physics, medicine and now and now a shortcut happens. Careful. This is something I did not understand the first time reading the paper. So the authors tell us now this was the theoretical process that would be great but we will go for a simpler hypothesis. So to say hey we are just a university we have to do something much simpler. We go and just say hey is our model of fivestep architecture cleaning routing planning grounding and execution really powerful enough on its own. So they say okay we don't do a supervised fine-tuning we do not do a reinforcement learning we just simulate a simple hypothesis and they say you know what we just have maybe only in context learning and prompting and we just tell here the agent here a sequential task and the beautiful thing is that they say their shortcut let's call it this way requires here zero task specific fine-tuning neither supervised fine-tuning nor anything about reinforcement learning So they say listen we don't want to invest the time and the money in this tuning procedure in the finetuning so therefore the training time is zero in our case we just want to prove that the methodology is working without actually any fine-tuning and I found this strange and I said why but okay let's go with their analysis so just to be clear now from now on they did not run any supervised tuning or any reinforcement finetuning for those experiments. So they say we take now any L&M great and we just have here a linear sequence. So their data agency they now present also with the data what I call a shortcut a reduction is now training free adapting here a serial shot to unseen data sets while sitting availing here some planning. So how do you do this? They say okay this is our linear pipeline we simulate in this simpler case no sequential modules 1 2 3 4 5 beautiful so this is now a program that orchestrate the different component so the first step that we do is here the cleaning the pre-processing step no this is now handled not by an LLM but a small rulebased controller in our simplified case so we do have a deterministic piece of code we say okay then the second step is here as you see the routing The routing is no done by an LLM GPD from beautiful and we have here one to three solution modes the classical linear the sparse model the neural model reinforcement learning trained or an LM based sequence generator and depending on where we route now we have in step three the planning and the tool calling now so for the classical you see this here with the red box here and neural it calls here a fixed tool chain so there's No reasoning, there's no discussion about it. This is a fixed tool chain to shape here a better feature set. Full stop. Or if we go maybe in LLM mode, it now generates a sequence of operations. So plan as a sequence and calls now the LLM define tools accordingly in a sequence. So you see what they do. They kind of shortcut simulate a simplified training scenario but they don't train it at all. They don't want to have here to collect the training data sets. They don't want to have the cost of the training process itself. And they say when our training free data agent wins then they say we can now confidentially claim that the victory was done due to the super superiority of the architecture of the five-step architecture that I showed you itself and not because we had a better fine-tuning data set or we had a training process or a longer training process or a better training process. It is just we simulate here the five-step process without any training which is a pity and okay but listen they get a performance and they augie now from the performance of their shortcut that this is the best model okay so let's follow this for a moment so they say now hey look at table three and we have here all the different beautiful benchmarks just look at the last lines if they go with a pure LLM a llama 3.1 a GPD system. Yeah, whatever you go. Never mind. Let's say we go here with the second column. The pure LLM has 85% accuracy. And now this simulated shortcuted data agent that is just simulating here the sequence has 86%. So they say look even if we just use an offthe-shelf LLM like a LM or a GPD system we are even one percentage point better with our five-step sequence operation than if we only do here a pure LLM which I would have said yes of course I mean you built here a complex agentic structure you simulated here and then you are only 1% better or if you look here the Third one 78% 78%. So the the success reporting by this group is very particular very special. They are saying we have a grand idea. We know what to do but we don't do it. We make a shortcut. We make a simplification. We have a training free methodology. We just want to give you the working proof if you want without any training cost without any training data accumulation. and we tell you, hey, yeah, it's it's not worse. At least it's the identical to a pure LLM, which I have to tell you, okay, maybe I'm not yet seeing here the real success, but okay, let's talk about the challenges ahead. I think this was just the very beginning here of a data agent. Sorry to be here super precise, but if you would go for the real stuff for the real data agent for genomics, if you really train it and you have really done a beautiful LLM that is optimized for a domain knowledge, careful because the human query, the human task will include a complex, deep specialized domain knowledge. This data agent is not just a simple data agent, but it has been built especially for the genomics task. And even if it can handle genomics, this would already be a miracle because normally you would have to subdivide genomics in different subcategories because the complexity might be too high for just the genomics in itself. So a data agent is so much more. It has to understand now which data to collect, which complexity data structure to collect, how to clean it, how to filter it, how to augment it, how to concatenate the data for a specific task for genomics for the human query. So there is not a lot of intelligence going into this particular agent just to find and collect the right data for the given task. And especially if you go to huge internet databases or a large data set, this can be really expensive. Now, because the agent has to try out to find here the data and excessive trial and error will be simply impractical computationally really expensive. Now, so future work must have some focus more about this in a later video to minimize the computational overhead. But how this is done, there's an interesting story coming up. And still we do have the blackbox dilemma. as our agent and the multitude of agent and have to communicate and now separated and are really only expert agents. No, as they become more autonomous and self-arning and self-dependent on other agents, you can imagine what happens. No, but also their reasoning path if you use your commercial systems are not really transparent. So therefore to ensure that the final analysis when you get the data that were generated by let's say a commercial data agent how can you make sure that they are trustworthy data that those data are the right data for your job. How can you make sure that the data that were given now to you that you pay for are also provable interpretable the right the cororic data. This is paramount here for the adoption here of data agent. Think about critical fields like medical anything connected to medical or finance. So a lot of challenges ahead in the future but interesting to see we build data for every little tiny little challenge. Is this the right way forward? And then an extreme training amount. And then we have all these agent talking to each other and you know as I've shown you in one of my last videos they have problems communicating with each other independent which protocol you use. So I think there's a lot of research in front of us. Don't dare to think that we have found any solution at all. We are just here trying to understand the complexity that we have to solve in the future because the current system are neither interpretable nor trustworthy to any amount. I hope you enjoy the video. See you in the next one.