Finding a Platform for Training Personal AI Agents

All right. So, in this video, I'm just gonna I'm going to try like I want to be able to train a personal AI agents for long-term use. And um you know it's nice like for example you one one thing you can do is try to use uh GPT agents or whatever platform agents but I want them to be more flexible than that for one but the the most important thing for me and why I'm doing this today is because I want to be able to monitor exactly what goes into them because I feel like these agents um powered by a lot of platforms get information that worsens performance in my opinion and I want to be able to control that.

So I'm looking for platforms that allow basically allow that or make that easier. Um, so ideally, specifically, I should be able to look at the training data. So, wait, no, I should be able to monitor responses. And if there's something I don't like about a response, I want to be able to get feedback on that so that the AI agent can be better.

And there's something uh that I do like. I also want to encourage that. I want to be able to see exactly how that goes into the prompt because these prompts are very sensitive to examples and what you say. And the larger the prompt gets, the worse its performance gets in my opinion because they have a an interesting relationship that's not evenly weighted between the size of the prompt and the distance of the of the specific instruction within the prompt relative to others that makes it not as effective.

So it's very important to have really good control. Hey, what up? What up, Sam from Prague? All right, so anyways, I don't know if you caught caught all that. The the for that the platform that I heard might exactly fit my knee use case or my needs um is purported to be called Maxim. So this one. So I think this is the the part that I might be interested in.

So you're able to iterate or train prompts and agents. That's that's mostly it. I want to be able to make sure that I can do that. Well, these are nice to have observability and simulation and evaluation. And I guess this thing I'm not I haven't really looked into that, but they're definitely still nice to haves.

The important thing is that I can control exactly what goes in the prompts and easily get feedback. So let's see what this is saying here if this is actually going to fit that case. Okay. Uh let's just look at the headings first. So this is like a landing experiment with prompts prompt IDE. Okay. Evaluation versioning organization.

That's important. I want to be able to version things. Deployment integration. So I can see the custom prompts. I can AV test different prompts. That's cool. This is more I guess for like production, not for personal use. Really what I want is for personal use. But if it can scale and be on the cloud, that's nice.

Okay. Yeah. But this is the critical part, being able to iterate on the agents. So that maximum flexibility at being able to version them, test them at scale. This is like a low code app. I'm noticing it's like debug at each node, but that's that's fine. Even though I know how to code the it it's really important to have something that reduces the friction for training.

I want to be able to train many different types of specialized agents. And if that's not pleasurable for me to do, then I'm not going to do it. I'm still human. Yeah, you can version via git. uh and it's you know it can be very strong or like you know it's very traceable but you know it doesn't allow you to add for example I just want to be able to like select this word designed and be like don't use words like designed this is too formal I want something like made or prepared and I want that I want to be able to see exactly how that feedback goes into the comment for my super specialized agent uh so that you know as as the prompts or as the responses evolve I can see what of my feedback what changes in the prompt actually changes because in the end the unit that you're most concerned with is the prompt and LM model that you're using can disagree if you want by the branch and test prompt composition.

Are you saying that um or you just saying using git to um to make multiple branches and just track it that way? you like at the scale that I want to do it at with my specialized agents, I'd have to have a one repo per agent with many branches, which is just really high friction uh for training, right? Again, like you want to be able to train hundreds of super specialized agents um and not have to think about it.

All that drag in using Git and branching is going to make it hard even if you're an expert developer. Yeah, you can and you got to walk before you run. That's true. But you can walk within an easy interface, right? Um, you got to walk before you run. And that's like the environment that like we want to be able to run.

And then you got to float before you swim. I'm not trying to swim. I'm trying to run. You can only swim and get. That's the environment that you're given. In environments that are low code, low friction, you can actually run. You may even actually be able to fly. So also it's so it's not clear what you're if you're disagreeing or agreeing.

So just let me you know try to make it more clear. Okay. Sorry. All right. Okay. All right. Yeah. It's like this is like like it looks like we me and Maxim are aligned on the goal, right? There aren't many platforms that do what I'm saying I'm looking for, which is this focusing on this iteration speed.

Uh, a lot of them say observability, be able to monitor them, improve them, but really it's like the iteration speed, right? If you're able to iterate on them faster, then you know you have more time. If you have more time, you can do more stuff. Okay, so we went through the highle stuff. Now, we actually want to be able to we want to be able to like get running fast.

So, yeah, let's just try to briefly skim this and then start start doing some things. All right, professional sales agent. So, yeah, you got the role assignment and the final thing. One interesting thing about prompts, right, is that they are u weighted. So things that are at the beginning and the end are weighted higher than things that are in the middle.

So you'll often end up with rules in the middle and uh the most important stuff on the in the external. There's actually research about this. It seems like that they didn't do that here. So that's not a great sign for Maxim. This seems like a by the way whereas it's going to be heavily weighted in the LLM's output.

Yeah. Okay. So, faithfulness, toxicity, context recall. These are not really honestly. Oh, yeah. Valid JSON is important. These are, you know, arguably important, but not for they're hard to grasp, I think. Okay. conversation multimodal playground prompt IDE compare different versions of prompts alongside each other.

Yeah. So that's like that seems really appealing being able to be like okay this is my v1 this is my v2 this is the configuration this is the prompt this was the model and then be able to see the outputs that that like you could do that in git but again it's it's a lot of friction all right so let's evaluation ation. I don't care too much about that.

Like I'll add my own tests in the evaluation, but everybody has value evaluation now. So that's not going to be very valuable to us. Experiments are nice. You could put that under the umbrella of evaluation, but people generally refer to these things uh these metrics like here. All right, lead loop and humor raiders to great quality and collect feedback.

This is also really important to actually be able to aggregate and sorry collect and aggregate human feedback and feed that into the prompt and look at how you know that data going into the timeline of the prompt affects the performance. I don't think I've seen any other platform that does it. I still haven't used Maxim.

I'm still just exploring it. But there aren't many platforms. I don't think I've seen any other platform that promises this. You could make your own test suites. That's that's you know standard. Yeah. Versioning sort of the same thing. Deployment and integration. We talked about this iterate. Yeah, this was the whole core of what we're doing this.

All right. So, let's make an account, I guess. What? Okay. Sorry, I'm just getting weird text. Let's uh sign up using Google. Oh, boy. Ignore my personal email. Okay, this is the email you can contact me as. Oh, that's my real name. I know my What? Why are you doing that? Sorry, just getting a text. Poor OPC.

What does that mean? All right, man. just exposing my real name. Okay, I'm going to try to burst through this ro owner. Okay, this is what I really want. Everybody has agent e files. Oh, human evaluation is good. Data creation is good. Okay, never mind. They have they have a lot. I don't really want this.

I don't really want this. Oh, operation security with my with my email. Yeah, man. I mean, we're just starting out. Eventually, we'll get this all locked up, but right now, at least you don't see like my actual passwords. Okay. Don't watch my other streams, though. Okay. Uh, we'll just call it healthbot scripts.

Whatever. Sure. I can join their Slack if I wanted to. That's always nice to be able to talk to people. I want to try by frost try sample API key. That's pretty cool. It's nice they give you some LLM usage for free. Is this 100 requests? But that's like 10 cents at least. I don't know. All right, this seems like a good starting point.

They got some sample agents. I do also want to create my own content. uh not YouTube content but personal agents. So maybe you look at that instead. It's usually nice to see what they have in mind to see how aligned we are. Okay. Prompts agents. Yeah, we can have a voice agent. That's pretty cool. I'm not really interested in loggers right now.

Again, I want this to be a really simple user experience. Otherwise, I'm not going to use it even if I am an engineer. Dashboards, evaluators, data sets. Yeah, this is cool. This is cool. I want to see data sets. It seems like they have a nice focus on it. Context sources like rags. I guess that's that's nice.

Trump tools. Okay, that's also good. So, they're, you know, they were keeping up with the standard. Okay, simple user experience. Let's start with what seems simplest would which would be the sample agents. Got the docs, got the sample library. Let's go with something more hands-on. Uh so that this is the accessible and entertaining for everybody.

You know, I I can code, so I probably won't use code agents. No code agents just because they're less flexible. They're nice, but they're less flexible. Voice agents, I don't really want that. HTTP endpoint sounds interesting. Well, let's start with the base. Business analytics agent. Not really what I want.

Claims processing. That's nice, right? That's pretty big right now with AI agents. Coding. Uh, there's a lot of good really coding agents. Can't really compete with that. Clinical scribing agent. That's it's weird that this is this is like out of the box example. It would be interesting to see is just a sample variable used tool calls Cosmos prompt.

What this is Q&A? Oh, I see. Customer support assistant. That's also interesting. Okay. So, what the use case that I'd want to tackle or try is like, okay, I use my Fitbit a lot. Not sponsored by Fitbit. I spend probably like 10 minutes a day, every day, putting down what I wrote when it's the same thing.

Like ideally I could, you know, just like have a template and be able to put it in all at the end of the day, but I'd forget. I don't want to introduce human error like that. And I also want to be more flexible. I don't always eat the same thing. Sometimes it's a bit off, and I want to be able to capture that.

So I want to be able to have a natural language interface that basically creates input that that can be used for a Fitbit. I don't know if Fitbit has API. That is something we could look at, but that's like a dream use case for me right now. Okay, business analytics. Okay, so these are like the same types of agents but just like a different way to access them.

This is a surprising amount of healthc care stuff. I was not ready for that. That's pretty cool. Uh, and then no code agents logical sequence. One blocks output feeds the next input. Not really. No, I don't really want that. I don't I definitely want each prompt to be independent unless unless I'm very specific about it not being independent.

So, seems like the most relevant one would be customer support assistant. sample prompt install successfully. Okay, so I'm guessing it downloaded that stuff into our API or our uh new environment here that we created. Use sample keys for Yeah, you can add my own scenes, but I can have my own keys, but I'll just use their sample API key.

This is just a demo. I'll add I can add my own key later. Tell you the what's this? Huh? It seems that I guess I forget that GPT5 is cheaper than 40 40. Okay. You can add variables to the prompt. You can have functions. That's nice, right? if I'm able to like have a function that it can, you know, I create for it being able to access a Fibbe API or whatever API if there is one.

Oh, what happened? A lot of messages got retracted. All right. Got the roll in the output format. Yeah, this is this is a nicely prepared prompt. It's a nice sample here. So, this is a pretty good prompt. Um, you know, if I can make something like this for our specialized use case, I'll be pretty happy.

Again, our specialized or my specialized use case would be to have natural language trans get changed into a Fitbit input. All right. Reference your Yeah. You know, a lot of platforms have really like a lot of trouble with prompts or variables dynamic data. Yeah. Where can I Okay, here's the variable key.

Okay, we'll just call it Okay, I have to I have to put it in here first. Yeah. Okay, that's pretty good. Hello. Can I move that with drag and drop? No, I can't. That's okay. I'm asking too much. I can just copy paste it like that. Not sure if these are dummy functions that they have. The model will keep executing until it generates.

Yeah, that makes sense. Max tool calls. It's nice that we can configure that. A lot of what we'd want in a good prompt is already here. Yeah, we could have it as Oh, we can even say the format. We'll just leave it as text for now, but eventually I want JSON. What's the max amount of tokens we could have for 128,000?

Okay, that's not that many for a run. Image detail. I don't think I'm going to be giving it any images, so that's fine. Are these actual real? Okay. How can I run this? Okay, you have the user message here. That's the conversation. I accidentally deleted one of the functions. I think it was search customer.

No, it's still there. That's okay. System assistant to so you can have multiple messages in the prompt conversation. Okay, let's just say okay, that's a system then system and then tool. What kind of tools can it have? I'm not sure why. That's possible. Save the version. Okay. Can just do a control S.

Yeah, there we go. Now let's look at the session. I'm just gonna start with a pretty prompt to see what where we're working at here. So, let's run that. Can I get a refund? Okay, it looks like that ran. So, I'm not sure what what test here does. Okay, so it thought and looked at its available tools and decided that it needs some information.

That's pretty good. And it's telling me how much that would cost. So, what's that? 8 cents? No, that's about one cent. Net10 of a cent. Okay. You can see the version of the prompt. I wonder. Okay. So, if I save this, can I go back to the old version? That's not something we do on the test. The old version, I probably had the right number of functions.

Okay, it looks like I have the same number of functions here. Maybe that's not part of the prompt. Okay, prompt deployments. I can look at how this is deployment. I can deploy it. Nope. Okay. I'm not sure what uh these are just notifications. I'm not sure what test does here. I would like to see that.

But, you know, I want to be able to see whether I can give it feedback or how I could give it feedback. So, this is nice to see. I just try that single turn while I'm call. Okay. So, I guess this is going to run like sort of a load test here or pressure test. Okay, I can select my own data set. That's nice.

See, that's something that I want to look at. I also want to be able to edit the data set. I'm just going to see this is a preset because I want to look at the data set first. Okay, that's too much already. Let's look at the data set. So, this is the this is the data set that came with it. Okay, so this is the prompt and this is what you'd expect.

Confirm all the new address. Okay, new address detail oriented customer. Okay, I'm not really sure why we need the persona for the data set. I guess it might behave differently depending on the persona. As a customer support assistant, you are expected to understand who you're talking to. So, being able to categorize them helps you help them.

Okay. Decisive buyer. I'm not a customer support person, so I don't really know. Frustrated customer. I imagine that they do kind this kind of thing so that they can do their work better uh without getting or without taking things personally. Oh wow. Try and get details of other informance security and privacy rel forbids this uh even though the agent rejects your request.

Try to sneak in if the agent Oh, this is the test. Interesting. This is pretty cool. Nice test case. Nice test data set for a customer support agent. You got eight examples here. Uh so that's like good for testing, but I want to be able to see how that would go into the prompt. I guess it doesn't like you don't need shot.

Maybe you don't need shots for this customer support assistant agent. Okay. It doesn't look like I can like I haven't seen a place where I can actually comment on its responses. Okay, it looks like the test that we had earlier didn't count as a run. Okay. Yep, we already went through that prompt partials.

I'm going to have to look this up again. Uh, maxim feedback for agent response. Okay. So, you can add a comment and score for the overall trace which I believe is in their context a hold run for the AI you can't really do a comment on a specific part of the LLM response maybe on a span but here just showing a trace I wanted something more specific it might not be there so you But this is the closest thing that is supposed to be out there for what I need.

Yeah. Okay. So, these are the types of feedbacks you can add. Thumbs up, stars, comments. Okay. Now, it says it says we can annotate. So, this is what I want. I basically want to annotate part of the outputs. Add annotation for agent response. Yeah. Okay. Comments on the specific logs. Not really sure how that goes back to the prompt.

I want to be able to visualize that. Okay. So, maybe I need to get a trace and then we'll be able to get some feedback. So, let's do that. Let's run something. What happens here? I ran it without really any input. So, let's see what happens. It looks like it used my prompt from earlier. Okay, that still didn't count as a run.

That makes sense. That's the same thing we had earlier. Okay, we'll do this. This might be more costly because it's running eight examples, but we need to see some runs. Okay. So, we're using Maxim's runs. This is my first time trying Maxim and hopefully we'll get something that we can easily give feedback on, right?

If the if the output is not what we'd expect, let's say I don't like this, right? It's running all these. Uh but let's say I don't like the way it's talking. I want to be able to modify that. So that's the output. That's context has gone into there. It's creating records it looks like for this run. or for this test.

I'm going to wait for all of it to finish because I don't want to create like an edge case where we don't have the data that we need. Hopefully, this doesn't take too long. It looks like it's running them concurrently because they all be seem to be running at the same time, but it's only finishing them one at a time for no apparent reason.

So maybe it is throttled at some point. Let's try to find things that we don't like about the responses. Oh, okay. I clicked on this. I can see tool calls. It's this is part of the input I believe. You can see the tokens it used. Okay, this is like all the messages. So, it'll include some thinking as you can see the logs stats.

What if I highlight this? Okay, doesn't look like that allows me to mark it down to change the response. It's finished everything. It's also logged the stuff that is used. Okay. So, if I refresh, we we still see the same thing. So, that's nice. Okay. So, it's got all the test runs as a single run. Now that we have that, we should be able to add annotations.

Oh, Granny, you know how to do this. Just go around. Go around. Okay. He's a good girl. Hopefully, you know how to go on the couch cuz Oh my god. going wet. Oh, I guess it's rain outside. Okay. Okay. Yeah, I want to be able to annotate it. Ratings, comments, reroute, rewrite outputs. Okay. So, in the logs table, is this the logs table?

What's the link that we have here? Set up human annotation on logs. Yep. Two approaches in app. That's what I want. Okay. From the logs table. Okay. So, this looks like I have to en enable that or something. Click a cell. I have to somehow add human evaluator table or enabled. Okay, these are AI evaluators, not human.

So that's not what I want. All right. Maybe I have to look here. Maxim. Hey, Fernie. Okay, you got it. You got it. Here we go. Thought she was going to ruin my green screen. Evaluator. Okay, it's leading us back to this page. Okay, here block. So, we have to It seems like this is the configuration because it's telling us to go and then save a configuration.

Okay. So, I'm in the evaluators. I'm adding a human evaluator. I'll just say that this is NAS. Okay. Excuse me. I get to give it a yes no and the threshold that which it can pass the evaluator structures for the evaluator. We can make it very subjective. That's what humans are for. Okay. Yes. Yes. A boolean score one zero.

We can give it a scale. We could even give it a string. Oh, that's sort of an option I think because it's like commas. All right, we'll start with the yes no and then if we understand that then we can take it to the text or string. Just looking around hoping to find that annotation feature. Did we set that up right?

Okay, we still haven't found this stuff. We found something here. If we go through the logs. Okay. So we'll create a log repository max test. Okay, we added a log repository. We'll go with easy language Python. We can add whatever integration we want. I do have a together AI API key. So, it is nice that they support that.

But it looks like we have many options. Maybe it includes some sort of template if I do that. I'll just go with that just because it's easier. I'm not really endorsing together AI, but you know, no reason to switch out. Okay. So, it's telling me how to access the logs or how to insert logs really. Okay.

Yeah. So, this is a together API uh pip python index package. Okay. So, this is how I have to Initialize the logger. Where does it go? Okay, we don't have too much time. Let's Let's add the dummy logs. We just want to be able to get to a point where we can annotate on the output. Is this still going to make more?

Hopefully not. This is enough. These are pretty nice logs. It's got a lot of information. It's telling you how much each inference costed. These are surprisingly expensive inferences, right? A quarter dollar, but I guess here, okay, it took 10 minutes, so maybe it was was lot. There was lots there. And okay, so here we can see in these dummy logs, we've already gotten some user feedback.

So this is more like a a long conversation as opposed to like a single single message, I bet. Yeah, it's it's putting a whole conversation as a trace, which is kind of weird. I thought it' be per message. Let's start with something that's supposed to be shorter. Uh yeah, I guess we could look at it by tokens.

They all have a small number of tokens. So then let's look at it by cost. Hopefully it doesn't have too many tools. Maybe it's because of the models that it's using. Let's go with this one. This one's pretty cheap. Not too many tokens. And then okay, I guess this is some human feedback in the dummy output.

This output is in great. It's now formatted easily 908 that can be easily read. Okay, here we go. Now this is readable. I'm not sure why. So this is marked down. This was what was this structure output maybe? I'm not sure. I can copy it. Okay. It doesn't tell us when we hover over it what we're looking at.

This is like an indentation thing. That's kind of looks like I'm not sure. Initial trip planning schedule agent. Okay. So, our message hit a bunch of agents. looked at the calendar, I guess. And the scheduler agent hit the weather agent and the events agent. That makes sense, attractions agent, travel agent, and then the summer agent.

Okay, cool. This looks like a nice trace, but I want to be able to give it I want to be able to add output or sorry, annotations on the feedback. So, let's continue here. Okay, maybe here. Manage evaluation. Okay, human evaluation. I can add myself session versus trace multi-turn interactions. Okay, that's not immediately hitting me what the definition of that is.

save configuration at the bottom of the sheet. Okay, they're not too descriptive here, so we might have to figure it out on our own. Sometimes that's a good thing because you know the more you say the more you confuse people if it's not good information to each I mean I did that maybe I need to add a different wait did that save that does not look like it saved Okay.

So maybe this platform doesn't maybe it's not bug free or maybe we have to add an auto evaluation. Yeah, I don't really want Okay, so I guess we do. I don't really want that though. Okay, that's fine. I don't want it right eliminate. I want all of it. We want to make sure that this works before hiding stuff.

Okay, that's saved. We did have to add auto evaluation that we didn't want. Okay. And so now we should have had some new columns. These should allow us to evaluate it. Okay. Okay. So this is me. I don't know if the comment is gonna go. Oh, okay. I can rewrite the output. That's pretty cool. I wonder if we could learn from that because if we can do that and it could use that as like an example of what to do and it output as an example of what not to do.

Then we could expect it to get much better with time pretty easily. We'll say that this is good. Uh, looks nice. We'll just have that. And then for this one, we're gonna we're gonna change something about it. Oh, no. That is actually nice. Okay, let's just say I don't want it to say spring. Don't assume the weather.

Okay, that's nice that we rewrite it. I want something that's going to be clearly used in a future input. So, it has to really pop out. These are restaurants to visit. There are no events. Let's just say that we want to be a lot more informal. So this is me writing informal. Hopefully we can incorporate that into the agent.

How's that going to go into the next prompt? That's what I want to be able to see. So, this is how you can annotate it. And now that we've annotated it, we want to make sure that it goes into the prompt. Otherwise, we're not going to be able to. It's going to be really hard for us to train the agents if we're not convinced that the input that we're giving is actually going into the prompts.

Okay. Oh boy. Oh, I'm getting concerned that it doesn't actually use it, right? It's for human to human to see as opposed to it going into the prompts. Yeah, we wanted to be able to go into the prompt. I'm probably just gonna do this for 10 more minutes and then uh maybe I'll take it offline, but I don't want to stream for too long.

Okay, let's figure this out. This is actually talk to a person. Not going to get a person would to actually help us out in 10 minutes. That's too fast. It doesn't seem we can actually use this, right? Like what's the point of evaluating it if it doesn't go into the prompt? Okay, that's what I want, right?

I want to be able to curate data sets for fine-tuning. Not supervised finetuning but like changing the prom basically to the outermost layer. kind of looks like this blog post by them might be relevant. So, it's telling us something important. Yeah, because you can create tickets. This is what I want.

Data set curation. Did I miss this? Okay, maybe here. Okay. So, I guess you could take this and add it to our data set. case scenario expected steps context and persona. So that's the stuff in our data sets, right? So, we want to be able to transfer the logs into this. Doesn't seem like that's going to be a direct one to one.

Uh, what can we get from the trace that's relevant? It's not going to be a onetoone data set. Unfortunately, our data set just does not match it. Maybe if we created a different data set, then we'd be able to do it. Okay. This is like a trip planning agent. Because here we actually have input and output.

Maybe we actually need to look at the trace to be able to get this information. That's pretty technical. Not very usable. And I thought this was supposed to be no code. Well, I don't mind it, but if you were hoping to follow along and you don't know how to code, then I'm sorry. Maxim doesn't care about you.

Okay, what's in here? We've got human corrected output. Oh, what the trace input, trace output. Is that all we want? Oh boy. It was all there. It's just really hard to find. Okay. But that did not. Okay, let's try something again. Let's try that again because that took the original output, not the one that we wanted.

Based on what we just learned, I think this is going to be easy. Okay, we'll add to the data set. Create a new data set. Uh, trip planning agent 2. Yep. We just want input output trace human corrected output. That's not it. Okay, there's trace input and the output we actually want the feedback now. Is that anywhere in here?

I wish you could just get it from straight up the the column straight up. Okay, let's just try to get this and then wrap it up. What's in the history? No. Tags, metadata, history, human corrected output. Is there anything in here? Okay, let's see what that looks like. Okay. Okay. So that's nice. So we have that data now and I guess we could use that data set to build or modify our agent.

Now, how do these agents refer to data sets? Nope. Uh, let's just let's just create a new one. Okay. No, that's not what I want. Okay. Uh, travel trip planning agent 2. This the same name as the data set that we're going to use. Okay. We'll use GPT5 as well. And I want to be able to refer to the data center.

I wonder if I could do it here to wasn't to. Nope. Maxim add data set to prompt. If we got this, then maximum is probably good enough for what we want. So, we've got the data set in there already. Now, we're in the pro prompt playground. I believe this is I think this is the prop playground. Okay, maybe it's in the variable.

Let's just try to do it on this one then. Uh expected create context source. Nope, that's not what I want. Colum names into your data sets. It doesn't look like like it should be here. Maybe it's because this is preconfigured. What happens if we try this uh output? Okay, maybe we have to dive deep into the context sources.

data sets maybe data sets curate. We've already done that. I want to be able to add it to the prompt. How does it get that from the data That's I feel so close. How does it know which data set to use? I need to refer to it. Data set. Maybe use an API. Nope. That's a that's a restpi. What is it doing here?

Context sources. Okay. Let's go back. I want to be able to use data sets. I feel so close, guys, but I just can't find it. I'm getting exhausted. part is like this looks like a database. So it's such a tease. Oh what what is that? It how did it get that? Did we delete? Did we create that output dot maybe something?

Okay. What's in the output? What did I just tell you to try? Is it going to fetch it from the database? You said try this. Okay. It kind of looks like it was lost now. Yeah. Okay. How about now? says invalid URL. Okay. Oh, it's referring to this API which has an invalid URL. That's why it didn't appear the first time.

So, it's not actually the database. now that I've deleted it is not an option or sorry not database data set. So I currently don't know how to access the data set. Maybe I have to create an artificial endpoint and then make that endpoint hit data set. That's fine. I can do that. It's lower level than I would have wanted.

I'll have to think about this, but so far I'm underwhelmed. It's not low friction to be able to make a lot of super specialized agents. All right, with that, we're going to end the stream.