Welcome to the Christmas episode of Agents of Data.
As you see, my colleagues, Sam and Joe, have really embraced the spirit of the season here.
Hats make me look a little scrooge-like, in comparison, but I'm not sure I could carry off the hats that they do.
(Laughing)
So, as we approach the end of the year, we're gonna take a little bit of time to reflect on what we think 2026 is gonna bring in terms of trends, particularly in the field of a agentic AI.
We've seen so much happening this year, so many developments and revolutionary advancements.
Predicting the future's probably pretty hard, but hopefully we can touch on some of the things we see coming.
So, perhaps, Sam, what are you most expecting to pop up as a trend, maybe in just the first half of the year?
Yeah, I think 2026 is gonna be the year of multi-agents and memory.
I think so, I think we'll see right now, we see multi-agent systems, but they're not very good, or they're hard to scale, hard to manage, that kind of thing.
I think we'll see far more multi-agent systems, so I think we'll see agents working for longer across a broader range of tasks, and then I think it'll be the year of memory, I think maybe even some of the big providers will start to build it in.
I think we'll see these models getting better over time on your own data.
Yeah, and does that mean you get, you see that as opportunities for the owner of the model, the users to actually train it the way they want, rather than training being all about the model provider itself gaining the benefit?
Yeah, I think we'll move away from these giant, massive parameter models that we see.
They'll still be there, because they'll be distilled down into smaller models or something like that, but we'll see smaller models that are better at cognitive reasoning and know when to go and look up information from a memory store, a memory graph, a knowledge graph, or something like that.
Yeah, so we'll probably come back to the memory in a minute, but Joe, what are your big predictions for the year?
I think there'll be something around AI fatigue.
(Laughs)
The pessimist over here.
2026.
No, I don't feel like people are going to wane off AI, or like the bubble, I mean, this bubble that we're ending 2025 in is supposedly gonna burst at some point.
I'm not talking about that.
I'm talking more about humans becoming, humans understanding more what they want from these AI systems and actually leveling up their, first of all, leveling up their prompts and leveling up what their expectations are of those systems to then actually start to see more of that innovation that's been promised in 2024, 2025, becoming more of a reality in 2026.
And what I mean by AI fatigue is at the moment, I'm seeing some users of these agentic systems simply put in a question and it doesn't bring back the response that you're wanting to see, first of all.
And that seems to be a sort of a hand-thrown moment of I can't use this tool or this tool doesn't do what I thought it was gonna do.
And I'm just gonna sort of give up and disregard it.
And what I actually think is gonna happen next year is people are gonna start to massively, drastically improve their knowledge of how best to prompt and how best to get out of these systems, what they're looking for.
And I think it goes back to what we were speaking about last time about how to run an AI native business has become a bit more focused on what you can expect to get from these agentic systems in terms of delivery.
So I'd like to see the humans improve in 2026.
That's actually what I'm most excited about.
(Laughing)
And hopefully being able to be a little bit of an equipped to judge what's good and what's bad and work out with it, make a judgment call on to whether what you're getting is fit for purpose, be realistic to your expectations, what you're asked for.
Yeah, yes, yeah, I think you're absolutely right.
Like we can all build a business in a day using a variety of different AI native products and applications, but like how much value does that bring and does that solve the problems that I was after solving?
So when you start actually then directing this, come out of this trough of disillusionment and into where we're gonna be, maybe we'll see humans actually start to use these systems a bit more effectively.
Yeah, I hope we get to the trough of disillusionment because the trouble is at the moment, there's so much noise about AI and so much, as I've heard recently, it termed AI washing and people stamping AI for every product because they've touched one tiny corner of it with a little bit of something egentic.
It makes it hard to cut through and actually get people to try things in earnest and really explore and just believe what you're talking about when you've actually got something to succeed.
So, Sam, you were talking about memory there.
It's probably a good one to pick up because people may have heard a bit about it.
It seems to be, it's becoming more of a thing people are realizing.
Maybe I'll start actually with Joe in terms of what you're seeing from businesses and customers that are talking about how much it's crossing their radar yet and what do you think that will grow?
Yeah, absolutely, it is.
Probably the number one question I get asked when we're deploying Maya, our agentic team into organizations is how do I get it to behave and think like one of my team members?
And for that, you need better memory, you need better ability to teach Maya or teach any agentic system what it is about your business that you would like being done or how you would like that product to be built, whatever that might be.
In our case, it's a data product.
So therefore, how do you increase the memory of those agentic systems to be able to handle the nuances of, oh, well, you know, this team calls that KPI X whereas this team calls that KPI Y.
It's all of that kind of knowledge that we're gonna start to see being onboarded at a greater scale into these agentic systems, I think in 2026.
I think it will also be used as a tool that just makes AI work better for your use case.
So you can imagine memories extend to things like episodes.
So like, you know, maybe your agent struggles at a particular task and by itself learns, oh, I failed here, I redid it this way and now it passed.
And now I will take that learning myself as the agent and not make that mistake again.
So, and that's what I was kind of saying that we'll see agents working over longer durations, providing more accurate results because they are building up this internal knowledge and memory of how to solve problems better.
That's, yeah, that's where I think the richest area for us to optimize and improve is setting up constructing scenarios in which our framework can experiment and learn and practice, just set up a bunch of homework the same way you would a human trainee, bunch of exercises to do and say make notes on what you're learning.
I think there's super rich areas for exploration.
And people have thought of training in this space a lot of being, I built a big training set that knowledge is absorbed into the model, the model itself, its weights are updated.
And that's something that an open AI on a topic can do.
But actually there's a lot we can do with memory to train the systems or you can do with the end consumer to train the system for your use case.
In terms of what types of memory, you could either, you could sound correct, you could actually drill into a little bit more some examples of what might be captured and how that differentiates.
Yeah, so I think like the most simple way is kind of long-term, short-term memories and feel like we've spoke a bit more about the long-term memories there.
So like the episodic memory, so like in our case that is our agenda guy, Maya learning from, as I explained, like things it did maybe slightly wrong and then self-corrected and also learning from patterns or things that users are doing with it.
And that's like long-term memory and then the short-term memory is kind of the stuff that I think all these agents already have, it's the conversation, it's the working memory.
So what it's currently doing and the entire conversation that is built up with the user.
But the long-term memory is definitely the place where I see 2026 being, that's where the focus is gonna be, like it's where the most value can come from I think.
And what types of things do you anticipate being valuable to capture as memories?
So we've spoke on a previous episode about the semantic layer for us.
So I see semantic knowledge as just another memory at the end of the day, a piece of information, linking business concepts to physical tables in a data warehouse and learning and building that up over time, form of memory episodes.
So yeah, when we've got things wrong, personal preferences, even like, maybe I want it to call me Sam, for example, I want it to greet me by my first name, it can save that as a memory, like all sorts of things, personalization, et cetera.
Joe, does that chime with the sort of things you're seeing people start trying to do?
Yeah, absolutely, like output styles is really key, like I think you mentioned there, like preferences, like yeah, I wanna see these agentic systems remember exactly how I wanna be presented with this information.
I've got a few gems and other things like gems that help do that, where I ask the same question and I always get the same type of response out and I'm actually quite happy with that, how I've trained it and how I've populated it, but actually wouldn't it be nice if these agentic systems were able to self-teach themselves that as they go through and pick up on the specific human behaviors.
Another thing that I'm really excited about for 2026 is seeing the productivity gains become a reality.
We've talked previously and made a few jokes around when we've automated all of our work, we'll just end up at the beach.
So first of all, I'm looking forward to getting to the beach next year and getting a bit of time to enjoy my automated work that I've got running in the back end, but in reality, we know that's not gonna be the case, it just means we're gonna be doing more work, putting more of those agents to work in that fashion.
So I'm looking forward to seeing some of the fruits of the groundwork that have been put in actually pay off in terms of, well, now I can definitely do two, three, five X and in some cases, some customers I'm working with, they're saying between 10 and 16 X more productivity.
And so then there's all of those additional, harder to reach tasks that I haven't been able to have enough time to do already, that now these agentic systems can actually help with those mundane operations that frees me up to start going over some of these to be a higher value ask.
So I'm looking forward to actually getting around to a few of them.
That opens up an area I can see us getting into in terms of measurement and evaluation.
The techniques of measuring how much productivity is gained, how effective the tools are in their space, probably is gonna get close to some of the techniques used by consultancy and management of humans and human processes.
Because we've come from a very machine-based mindset, but really they're enabling processes that you'd look at in terms of performance optimization or measurement of system efficiency.
One thing if I could add into a wish list, by the way, sorry, just that you've got my brain where now, one thing I'd add in is the, I hate it when agentic systems feel like they finished and they're prompting me, like Chachafiti, Anthropic, they're all, they're genuinely, they all do it, they all say, oh, now you've done this, why don't you go and do this instead?
Or why don't you, or his X, Y, or Z that we can go and do now, or would you like to finish this task?
And more often than not, especially with, I don't know if it's just how I've got my Chachafiti set up, but more often than not, I've not finished with that task.
And I actually, I wanna dive into it in a bit more detail and apply a bit more human thought.
I'd like to see these agentic systems get a bit smarter as to how they understand when I'm ready as a human to move on from a particular episode or a particular conversation.
No, I think that's probably something, actually, that is exciting.
I mean, we look at the model advancements that are coming out and like, I mean, we're gonna, as we have seen this year, we're gonna see GPT-6.
I mean, there's even rumors of that before the end of the year, not sure.
But we're gonna have new anthropic models that are just gonna get better.
And it's just like me as a software engineer building on these things.
Like, it's a weird feeling when you know that like you see the anthropic model drop.
And like, the first thing I wanna do is toggle that on for Maya, run our evaluations, run our benchmarks and see it.
It's almost like I'm giddy, a bit like Christmas.
You've got the model comes out and I'm like, how much better?
I'm on a packet, yeah.
How cool is this gonna be?
So I'm quite excited to see the advancements next year because it's been pretty crazy.
People start to see not just like we've got LLMs now, but like diffusion models and different models coming out.
What's the diffusion model?
I mean, Julian.
I'd say brilliant.
I think the area of active distillation, model distillation is probably one of the ones I'd see a big opportunity for us to get more involved with.
As we get to the point where the models are more advanced and model distillation involves using a large language model to train a small language model for a specific task.
And I think there's a lot of spaces within any agenda framework which we get a bit better at understanding it.
We put the models in the framework in a box.
We know it's giving it inputs and outputs to set things up, say, okay, train a specialist model for this particular piece of the system, train a specialist model for that piece of the system.
And hopefully them being smaller, there should be less memory hungry, less compute hungry and also faster.
I think the key is the tooling like some of the big providers like AWS are putting around doing model distillation is getting way easier.
It's stopping being a task for a specialist high coder in a Python notebook and more, which is another thing you call from AWS or GCP.
So the simplest way I can describe the diffusion model, by the way, is if you were to watch chat GPT generate tokens, it's one after the other.
The diffusion model, it just all arrives.
Like it's like one--
Oh, like how images do?
They're still actually-- Are they still--
They're still token by token.
Like the image models, but diffusion models essentially, like all of the tokens appear at the same time.
So it's like, if you were writing some code of a diffusion model, you can go and watch this on Google.
I've got a diffusion model.
Like the tokens don't stream out one, like sort of one by one.
You'll just see the code kind of like, Nice.
I'd love to, I wish, I don't know the full technical details of the difference between diffusion and LN under the hood.
So I won't try and explain that now, maybe another time.
Yeah, maybe one more time.
It's showing us it puts a parallelization effectively.
But essentially the speed of them is insane, but the accuracy is not as good as a foundation or a state of the art LLM right now.
Wow.
But yeah, some crazy stuff.
What about quantum computing?
We're gonna see that with LLMs next year.
I'm gonna pass on that one.
Or crypto and the blockchain.
(Laughing)
Yeah, I think that's, I'll just get the magic eight ball out for those.
I don't think this.
To be fair, just aside, there's something really interesting I've seen.
I don't know if anyone's probably had X and socials and stuff, have you seen that they are, they've basically put all of the state of the art models against each other doing cryptocurrency trading?
So I can't remember the exact website, but it's really fascinating to watch like DeepSeek, these Claude 4.5, these GBT5 high, all trading cryptos and how much money they're making and the risks they're taking.
I definitely recommend people checking out.
It's actually just really fascinating.
Have you seen the bioengineering stuff?
Have you seen that?
This is really exciting.
I can't wait to get to run a genteck system off an actual human brain.
You know what I'm saying?
Cell cells grown in labs in Switzerland where they've got like a donor, like an ethically approved donor of human brain cells that they're now growing.
A, it's called wetware, right?
And they're growing cells and then working out, A, how to sustain the human life of the cells, but B, also how to then use them to create systems.
So I'm excited about that.
It's quite a frontier model.
Yeah, it's still quite a frontier at the moment.
The other fascinating thing coming will be world models.
So essentially people who think we need AGI and stuff like that don't like, LLMs are actually quite dumb.
I know they seem to everyone, they seem like magic.
They're really not, they're just auto complete or prediction engines.
World models are actually like, you can see, again, I think it's like a Google Deep Mind are working on these and it's like a model that like truly understands the world.
So they're starting to build like, you know, 10 minutes of a video game that as you're in that video game driving around, the world is generating around you and is completely bespoke to your game experience.
But it's so expensive and they can only generate like a few minutes at the moment.
But this idea that the models and some technology, I don't think it's LLMs, it's something bigger, can actually generate a world like, and understand physics and stuff like that.
We see it in LLMs, but really they're guessing based on pattern matching at the moment.
I think like you often see the models come out and one of the things they do is they plot a ball dropping and it's like a shape moving around in a circle and how it reacts to the physics determines how good the model is, but really it's all a bit of guesswork.
That's the extreme end, but the definite train we see now that's turning over this year of the capacity of what can be done within a task or a problem across this and going into a system has grown significantly in terms of how much autonomy you can give the model or the group, increasingly the group of models.
And I think having groups working in a fret and gently framework is only gonna increase the trend in terms of specialized roles and specialized checks on what's being done.
But we've seen a strong trend even internally that previously at the start of the year, we were expecting very much speed to be the priority of just like, okay, have it return, it does a simple task well, return to the user quickly so it can do the next task.
And we thought we were gonna have to optimize on speed, but what we saw over the year was as the capacity of the models got better and we got better at building the system, you could give my own much more complex task and through that, then it becomes more asynchronous.
You can leave it for longer, speed was less important, actually accuracy became the priority.
So you could trust it with quite a complicated instead of instructions, leave it running.
It doesn't really matter if it takes five or 10 minutes so much, if you can do it and leave it with the whole task that you would take a human an entire day.
And I can see that trend only increasing the complexity of the task that's being tackled grows.
And I think like the other thing that complements that is you can see the agentic platforms moving more towards being like task management solutions.
So essentially you have, you definitely always wanna, in my opinion, go for quality over speed most of the time.
But the quality and the speed being lower with the quality being higher, if you essentially take 50 tasks, 50 is a large number, maybe 10 tasks, and then you're actually saying go and do all 10 at once, or materially or way faster than you were, even if each of those tasks still seems a bit slow to you, even though it isn't, it's probably still faster than you could do it, but you're essentially now doing 10 or 20 of them at once because you're able to go into some sort of UI and just drag in a Jira ticket or write a load of tasks and then--
The barrier for that, the challenge is how you measure the success or how much you can automate the checking whether it's done a good job or not.
May well be where you get the opportunity to bring in more of the AI judge capabilities to do the first level checks before it comes back to the human.
So a lot of exciting things.
Any of those, can you catch your fancy, Joe?
All of them, I mean, yeah, you raised a really good point, Jim, in the start of the year, we went to speed, and really now we're becoming more task-driven at the end of 2025.
And I think for me, it's about how do we get out of the trough and how do we get more of the leaders in the business adopting those egenic first-light mindsets, and there was something we've already spoken about previously, but also how do you actually adopt some of these technologies in a way that has got the proven track record that's got the tests and you've got all of the evidence in your back pocket.
And I think now that we've seen a year of people playing around with some of these different ways of doing things, we're gonna start to see some of the technologies emerge that have actually got the rubber stamp approval of, yeah, this thing, this egenic system actually works.
By the way, we've got one, it's called Maya.
It's an egenic team, it makes you 60 and X more productive.
But crucially, we'll start to see adoption of some of these systems that have got those proven track records, and we're looking forward to seeing the gains that organizations will find with that.
I do think we already see it, we get it from Maya, like the customer store is coming out.
The success that people are having with things will become way more prevalent.
I mean, you already see case studies from Anthropik or OpenAI, you see we've got our own case studies of customers using Maya.
We're gonna see more and more success.
Yeah, and I think what I'm hoping for is a little bit more maturity in the market will help people be better equipped to judge what's working, what isn't, and what works for them and what they're willing to accept.
Because at the moment, you've got a mix of the classic early technology adopters want to dive in, and then there's maybe some more risk-averse potential that customers are avoiding it as it becomes a little bit more commonplace.
And maybe actually a bubble bursting wouldn't be too bad because clearing out some of the hype and as we say, some of the AI-washed products might not hurt because it's hard to cut through with a message when you've got something that really works.
And a lot of, I think there's crazy money going in and crazy seed rounds, lots of unproven companies of young Silicon Valley hotshots or whatever, and it's like, at some point, there's gonna be some consolidation or something going on.
There's just too much money around at the moment, I think.
Too much for us.
Yeah, I mean, we're seeing it with cloud data warehouses as well, and there's something like 300 cloud data warehouses across enterprises today.
And there's a few of them that are household names and a few of them that are coming from the big providers as well, but we're starting to see storage and compute become more and more of a commodity and more and more segregated.
And in our world of data engineering, I think that's only an exciting thing to see more modularity, more choice to the customers based on, essentially, pure economics.
So that'll be exciting to see where that goes and to see how some of those big players react and respond in that space as well.
But fundamentally, to your point, Julian, we're gonna see some wheat cut from the chaff in terms of the agentic models that are there today in out in the world, out in production, that actually aren't necessarily agentic.
They are AI washing, and they're just there as a bolt onto ticker box that says, yeah, we do AI.
And we're gonna start to see that.
I think you'll probably see that in a range of products where there'll be differentiations between, at the moment, a lot of companies are talking about AI and only a small number of delivering.
I believe you'll start to see the proportion of that amount they talk about AI dropping back if they haven't really embraced it or made it work, or it no longer becomes the lead feature that gets demoed when you haven't got proven results.
I mean, we're seeing that already with some of the consolidation of the tech companies in our space, recent mergers and acquisitions.
If you look at the course of 2025, especially in data engineering and data integration platforms, the field that we were playing in at the start of the year has now, most of those big players have either been acquired or have merged.
And so we're starting to see full maturity curves for entire segments of the Magic Quadrant, where it then becomes, well, actually, the only players that are still in this space going into 2026 and beyond are the ones that either have an AI native product, like Nia that's out of the box, an agenda team to help you, or other players in that space that have got a sort of a legacy stranglehold.
And so then I'm quite excited to see what happens to some of these new up and coming AI native tools that are in there ready to disrupt in these sort of existing ways.
That's just the way in which we do work.
That's just the way in which we do process X.
Well, you know, if 2025 is towards anything, it's that that is just not gonna, that status quo, maintaining that status quo is not gonna serve you as a business leader.
Well, it's not gonna serve your tooling, your choice of tooling world either in terms of getting those productivity gains.
So I have a question.
If there was one thing that AI doesn't solve for you right now, but you want it to solve for you next year, what would it be, Julian?
Yeah, so I think proactive investigation of data, the ability to say these are my care abouts, keep watching it, keep monitoring for this and be a move from a pull-style model to a push-style model.
So the pull-style model is I ask for a request, I ask for analysis, I ask for transformation, and I get pretty good results.
But I can't, that's limited by my imagination and what to ask and my capacity to think, to remember, to raise those questions.
I'd really rather give my AI assistant a brief to say, okay, here's what I care about, here's what is most important for me and my organization, my business.
Can you keep probing into the data, keep looking and then alert me or reach out to me when you found something interesting?
Sounds good.
What about you, Julian?
I would like that in my personal life, as in I've got my calendar, I've got my social groups, I've got my messaging, it would be nice to have a personal AI assistant that's pushing a summary of what's happened across the spheres of humans that I care about so that I can be more proactive with social events or with X, Y, and Z.
So what you're describing there in a push model, that's actually really exciting, I can see that.
I'm excited by that.
I would like all of my AI systems to know what date and time it is.
(Laughing)
At the moment, if I'm asking it to do research or asking it to do forecasting or X, Y, or Z, I've always got to tell it, no, no, no, no, today is December 2025.
It's not June or it's not when your model context Monday shut off was.
And so yeah, I'd like them to have more, I think you were saying at the start about more sort of like just that world view modeling, how do worlds operate?
Like if I ask Chacobuttee to help me prepare my Christmas day cooking plan, and if you give it a very specific set of, these are the things that I would like to present on the table, I've only got one oven.
Go and work out like the cooking order.
It still manages to produce like a really great plan, but it forgets the fundamental assumption I've only got two trays to put into the oven.
And so just getting that sort of like general intelligence level up, and also the ability to have it think like humans think about some of these problems, I think it's gonna make all of these AI models improve over time.
So what would you prefer to happen there?
You'd refer Chacobuttee to come back and say, Joe, unfortunately you aren't going to fit this in the oven, but unfortunately our lines aren't very good at saying no.
I want Chacobuttee to message me a week before saying have you ordered your dad's vegan meatloaf yet?
Wow.
And so that I don't forget, right?
And I want it to understand that I wanted to remember that my dad's vegan, I wanted to remember that I'm gonna forget about that.
And I want it to encourage and almost, I can't enjoy behavioral science, so I want some positive nudging, some positive, helpful insights to be generated ahead of time.
Perhaps my next Christmas will have something that's gonna remind me to order my vegan meatloaf.
Chacobuttee might start serving some ads soon, it might just say here is a bigger oven for you to buy.
Yeah, but it's forgotten the parameters of my kitchen.
I've got a very thick space.
It doesn't know what it doesn't know though.
Exactly, yeah.
So what's on your Christmas wish list?
Well, I asked this question because I knew it was really tough.
(Laughing)
So I don't know, I mean, what propped into my head then as we were talking about yours, Joe, it was probably that I just want the LN to tell me when I'm wrong, instead of assuming that I'm always right.
Oh yeah.
There's nothing worse and actually more unproductive than it running away with something that I've told it, which is fundamentally not right.
And the thing where it's like, "Oh, great point, Sam."
Yeah, and I'm like, "Well, it's not a great point though, is it?"
It's just something I'm doing.
And I just don't think that's a fundamental flaw in LN, so I'm probably not gonna get, they'll get better as they ground on and they reason and they get better at reasoning, but there's still just a fundamental problem in text prediction, so that's not gonna solve.
So you almost want a positive or negative, you always want a sentiment dial.
Not a temperature.
I just want-- Yeah, you didn't care about it saying, "Yeah, yeah, yeah, but it's more about, no, it's more of the positive and negative reinforcement that provides you with it."
I'm probably just asking for AGI at the end of the day.
(Laughing)
I suspect that making--
You want that Christmas joke.
Making the model say no in the right scenarios is probably gonna be AGI level at this point.
But then the jokes aside, I don't know, for me, what's most difficult for me right now, the software engineering stuff, the coding agents are getting so good, but they're still, I just think this is fundamentally a really hard, maybe less so technology problem in the sense that the other ends aren't good enough, but being able to work across multi-service complicated problems, and I would say, as a software engineer, we deploy many services, and the change I need to make my cascade through a few deployed services, and at the moment, they're just not good at applying those or testing those all, and I think that's just because it's a fundamentally very difficult thing to do that humans are really good at, because we essentially have a computer and the ability to reason and go and look around and change our minds and elements.
Certainly, getting better or having more robust techniques and evaluating that the task has been completed correctly and sensible checks for reasoning and common sense applications about has it caused any knock-on problems, is key to us being able to work with them in an asynchronous task-based way.
If we get better, a better way to judge success is the next thing we've got to unlock.
This is the multi-agent stuff for me, though.
I don't think, part of it will be limitations in NLM in the context they have.
Part of it will be, it's just a software engineering problem that we haven't made very good complex multi-agent systems, and working across a very complicated code-based multi-deployed software system requires lots of different agents with different understandings of different domains.
You kind of want your SRE agent, your Infosec agent.
Exactly, and I think that's popping up.
I think as we see, I think we spoke about the A2A stuff before, and we'll see more agents talking to more agents to solve problems, but I still think that's six months a year off being really useful.
Yeah, there's definitely room to explore, though, in that scope of agents with different roles talking to one another, and actually showing how talking through a problem often solves it better than trying to reason it through, as a single individual seems to be true for agents as well as humans.
So, carrying on the spirit of seasonal questions, what's your New Year's resolution related to AI and agents, then, Joe?
Oh, oh, that's a tough one.
Okay, yeah, so one thing I learned last year, it was don't go into a new tool or a new technology, assuming you know its limitations.
And for me, that was a really big, steep learning curve, and I think it's one that I'm helping our customers get on board with as their job adopters, gentec team members like mine, into their organization, is don't assume that it can't do something, or better than that, rather than a negative, for your positive wish list, Sam, you know, ask the agentic system how best you can utilize it.
Ask it, how can I utilize you best?
Give me some examples, give me some case studies.
You'll be really surprised with what it comes back with.
And if it's a good, decent agentic AI framework, then it will be able to answer that question with a good level of rigor and depth and understanding of what its purpose is, what it's there to help and end user to.
And so for my analysis and exploration of AI frameworks, agentic products next year, my newest resolution is always gonna be interrogate the agentic system in front of you before you start using it, so that you don't get shoehorned into focused on what you think it should do, and let it tell you what it can do to help you.
That'll be my newest resolution, and it's a bit abstract.
Sorry, yeah, very lengthy as well.
Apologies, yeah, I'll add to that.
I will say respond using fewer words when I'm podcasting with Sam.
(Laughing)
Go on, let's go on.
Yeah, short and sweet, I think.
(Laughing) Just make sure we won't run out of time today.
But yeah, so I probably have become too biased towards a few tools that I've really learned to get the most out of, and I feel like I could carve out a little bit more of my time to explore some more experimental software agents or general agents.
There's some out there on my list that I haven't got around to yet, so trying to do more of that.
And what's on your resolution, Jamie?
I think mine is probably, before I start doing something, it's actually decide my current success criteria from judging whether it's succeeded, the system succeeded, or failing for the task I wanted to do.
Go in there with actually set up, set up a way at which you could fail.
Plan the metrics out first.
Yeah, we've seen that a lot with our customers.
That's kind of my view.
And I guess my final wrap-up question is, where do you hope we're gonna be by Christmas next year?
Bob Aydus.
We'll be on the beach.
Automated, 2027.
Yeah, should be fully.
Hopefully we'll be fully automated by then, and our bots will be doing this podcast episode this time next year.
I'll be managing them from my companion app mobile app while I'm sat there with a nice mojito.
My, excellent.
Thank you for your meat left.
So thank you, thank you very much for joining our Agents of Data podcast in this festive season.
I hope you enjoyed this, and tune in again next time.
Cheers guys.
Nice one, Jim.
Good question.
Good seasonal