Sign in or Join the community to continue
Why Are People Excited About Jev?
Posted Nov 06, 2026 | Views 10
Share
Speakers
Erica Hughberg
Community Advocate @ Tetrate
Erica Hughberg is a technical leader and community advocate passionate about helping engineering teams build scalable, secure, and human-centric application platforms. With a background in software engineering and a deep understanding of cloud-native technologies, she specializes in driving the adoption of open-source projects like Envoy Gateway, Istio, and Kubernetes Gateway API, which enable organizations to simplify traffic management, security, and API distribution.
As a maintainer of Envoy AI Gateway, she plays a key role in shaping the future of API infrastructure. She focuses on features to ensure organizations can securely and efficiently integrate AI-powered services while simplifying traffic management, security, and API distribution. In the Envoy community, she drives collaboration, mentorship, and contributions that advance the project and its adoption.
Lastly, as a believer in the power of storytelling, Erica enjoys translating complex technical concepts into engaging, accessible narratives in the form of social media posts, conference talks, podcasts, and educational content.
+ Read More
Demetrios Brinkmann
Chief Happiness Engineer @ MLOps Community
At the moment Demetrios is immersing himself in Machine Learning by interviewing experts from around the world in the weekly MLOps.community meetups. Demetrios is constantly learning and engaging in new activities to get uncomfortable and learn from his mistakes. He tries to bring creativity into every aspect of his life, whether that be analyzing the best paths forward, overcoming obstacles, or building lego houses with his daughter.
+ Read More
SUMMARY
What happens when you keep a bear in a cage built for a bunny? That’s our analogy for AI agent security: the agents aren’t the problem, the cages are. OAuth was designed to stop people handing their passwords to third parties, not to fence in agents.
In this catch-up episode of Debug Mode, one of us is just back from Japan and from AgentCon and MCPCon EU in Amsterdam. We get into why loading more skills can make an agent worse, and what is actually new about Jev, the fast classifier API, when classifiers have been around for years. Think support bots that bring in a human before the customer loses patience, or a math tutor that can quickly tell whether a seven-year-old is still asking about math. Then we take on a question from Amsterdam: is computer use just a stopgap until everything goes headless?
Which service do you think will never ship an MCP server? Tell us in the comments, and subscribe for the next episode.
+ Read More
CONTENT & TRANSCRIPT
Erica (00:00) When people are afraid that AR agents are going rogue, I feel well, I'm afraid that the humans are not spending time thinking about how to build a better cage.
Demetrios (00:11) Dealing with just the chat feels very boring.
Just march down files.
Like really, that's the best we've got.
We've gone backwards.
Erica (00:20) So we built machines to work alongside humans, not instead of, but alongside humans.
Demetrios (00:39) There we go, we're back.
Oh my God, it's been like two weeks.
Erica (00:45) Two weeks, forever, you know.
Give or take, about the same amount of time.
So I feel like it's...
Demetrios (00:53) That's like almost a year in AI years.
It's basically a lifetime, so...
Erica (01:00) Yeah.
So we clearly have a lot to catch up on.
You've been like all over the world.
You've been to Japan, you've been to Amsterdam, which are probably two places that are quite different, but also I would suspect have some I never been to Japan, but in my mind there is like I believe Dutch people and Japanese people both like order and structure.
So I don't know.
Demetrios (01:29) One thing that is wild about Japan, though was so I I don't know if I can go that far about the order and structure piece, but what is wild about Japan is that everything is super clean.
So like the streets are very clean, yet you find no trash cans anywhere.
There's no public trash cans.
So you have to kind of do do what you do when you go camping where you like pack in the city.
Erica (01:55) If you are a responsible camper, that is, if you are a good camper.
Demetrios (02:00) Yeah, you pack in and you leave no trash behind.
And so I found myself walking through the streets of Tokyo with like coffee cups in my bag.
Paper coffee cups, not like a coffee cup that I got as swag at the event.
No.
Paper coffee cups that I got from that morning, and I couldn't find a trash.
Erica (02:26) How many coffee Okay.
Dude, I think you just like revealed a different problem here.
So you're saying multiple coffee cups.
I think that that to me...
Demetrios (02:38) I should recycle reduce, reuse, recycle.
Is that what you're trying to say?
Erica (02:41) I'm just more concerned about how much coffee you're drinking.
That alarmed me.
It's like it's not like oh I've been carrying around coffee cups from the last few days.
No, no, no.
This morning, you said.
So I'm like, Yeah.
You know how they said that millennials can't buy houses because they buy too much coffee.
Maybe it was people like you they were speaking about who
Demetrios (03:06) Well, not only that, I will tell you also.
Japan has an incredible coffee culture.
So I had to sample the forbidden fruit.
That's one thing.
The other thing is I'm not buying a house because I spend so much damn money on this denim jacket that I've got.
I don't know if you need to.
Erica (03:26) It looks amazing.
I think it looks great.
I don't even see the full one.
I don't I don't even see the top of it, but it does look great.
Yes. And uh
Demetrios (03:37) After I you know what it made me feel like you ever you know that meme where it's the people dancing on the dance floor and then somebody on the side being like oh, but they don't know X, Y, Z, whatever it may be.
Erica (03:48) Dude, no, I do not know this meme.
I feel uneducated, like I am not part of internet culture.
But okay.
Demetrios (03:55) Well, so it's like stick figures and they're dancing, having a great time at a party.
And then there's somebody in the corner, and the meme is that you put something there where it's like, Oh, well, these people don't know about XYZ.
So even though they're having a great time, I've got this secret that they don't know about.
and I felt like that meme as soon as I got this denim from Japan because now
like jokes on you folks you don't know how cool I am because I've got denim from Japan you know
Erica (04:31) Okay, but did you buy converse shoes in Japan
Demetrios (04:37) no and I also didn't get the um the tiger
you know the Osaka tigers I did go to Osaka though and I saw a bunch of cool stuff I won with my daughter a gigantic oversized hello kitty from those claw games.
Oh and those are an absolute scam.
I don't think anybody actually wins, but okay with it because out of the twenty dollars that I spent, we won once. So I...
Erica (05:07) Okay, how big is it?
Demetrios (05:09) It's huge.
It actually didn't fit in our luggage, so we had to carry it on the airplane as like a separate piece of luggage.
Erica (05:07) Carry on. Yeah. Yeah. Carry on Kitty.
Demetrios (05:22) Yeah, that's the debrief of Japan, I guess you could say.
There's so many other things, and I can see why people fall in love with Japan.
You know how it's one of those things where it's almost like Bourgeois, I'm cultured, I speak Japanese, I know about the Japanese culture, especially in the tech scenes.
You know, somebody's Oh yeah.
Oh, they're so cool if they speak Japanese.
Erica (05:46) Well, I it was really trendy at one point to like learn Japanese.
In my school in Sweden, you could pick Japanese.
This is like quite a while ago now.
You could pick so in Sweden, English is your like second language to learn, then you pick a third.
Like so you pick, or the second, like it's your first second language being.
So then you pick French or German or Spanish, and then you get a chance when you're 16, you can pick another one.
And it was very cool because they offered Japanese as one of these languages, right?
And who was very excited, or I say who, not a particular who, but more the general who.
It was clearly everyone who was obsessed with anime.
And that it's a fantastic overlap of the people who are in tech and the people who like anime.
So hence who learns Japanese, right?
Um I was lazy, I just continued with German because you know, just just do what you know and try and maintain those grades.
I did it.
Demetrios (06:50) German such a beautiful language, too.
We just got a hat tip to the linguists who dive deep into the German language and really can appreciate all of the sounds that it can make.
It's a brutal one.
Erica (07:07) I think like German has this reputation of like being this like not beautiful language.
But I think it has like
Demetrios (07:18) No, really?
Okay, well, all I will say to that is maybe German from Austria can sound a little nicer on the ears, but German from Switzerland.
I swear to God, that is the worst type of German there is.
And I will hands down, it's like that other meme, like Change My Mind.
Please change my mind.
Tell me that it's not go listen to the radio in Switzerland in Zurich and tell me that's not the harshest type of German that there is.
Erica (07:50) And I was going to tie this wonderful thing into the world of AI because the Gates Foundation, right, is now committed to make AI available in more languages.
So people can talk to them in whatever beautiful or non-beautiful language they natively speak.
Demetrios (08:07) Well I'm glad you brought it back around because we've been pontificating about everything except for AI up until now.
And while I was at these different events in Japan and Amsterdam, uh I'll go over Japan first.
Had a blast talking to a few people.
I got to catch up with Ola from she created and is like one of the lead maintainers of the skill.
Erica (08:33) MCP skill, yes.
Okay, dude, right.
That is like I'm so happy.
I even messaged Ola about this because I was like, I'm so excited because this solves a problem.
I've actually been trying to solve what me and my colleague have been out trying to solve this trying to figure out what to do and then I was like I saw okay
Demetrios (08:07) How are you using it? Wait what which part of it are you try were you that...
Erica (08:58) Yeah, we were trying to s yeah we were trying to solve the
how we could instead like being able to consume skills on the flight over MCP and being able to ensure that any consumer of said skills always had the latest up to date version of the skill.
And so I was so excited to see the skills over MCP because one of my most f biggest frustrations about skills is keeping them updated and being able to Yeah, and I think having this better discoverability of skill I think skills over MCP is like the solution to this.
So I'm super excited.
Demetrios (09:44) Well, that was a surprise to me because I had understood skills over MCP as something that was oh, cool.
Here's a skill that will teach you how to better use this MCP server.
She was like, Yeah, I mean, that's one way that you can look at it, but you can also look at it as what you're talking about.
And uh she said that it was more common of a use case, what you were saying as a distribution mechanism for different skills.
And so that was surprising to me because I didn't think about that at all.
Erica (10:18) I think it's like think about all of these like vendor products and just think about big vendors, like if they're big cloud providers, right?
Imagine like AWS, Asher, or JCP, like you could what if they had like skills over MCP for how to interact with all of their
services and you can like pick up the right skills at the right time and you don't have to worry about like maybe maintaining updating those skills yourself because Asher or you know Microsoft, Go, uh Amazon, we'll keep those updates.
I think that to me is very exciting.
That you you don't have to go in look on the internet.
You can just literally okay, here here are these like skill providers almost that I want to subscribe to.
Demetrios (11:08) Well, yeah, it's funny you mentioned that because I literally was talking with James Ward two days ago, and he's working at AWS and he's working on agent experience.
And so one of the things that he touches really heavily is the AWS MCP server.
And we went deep on hey, how should the MCP server be?
Should it be like one gigantic server that can touch all 16,000 of the AWS services, or should you have a specific Dynamo DB MCP server and then a DuckDB MCP server, and all of that.
And so he was saying that he's actively trying to figure that out.
It doesn't feel like the small MCP servers are the right pattern just because you've got so many that then AWS now has to maintain 16,000 different MCP servers and skills over MCP is one way, but the problem with skills in general, in my mind, are that A, you have to optimize the skill to the model.
Yeah.
And you really have to there's an art to creating skills.
So if you are supporting skills over MCP as a provider like AWS, the matrix gets really big really fast because of how many different skills you have correlated to how many services you have, but then with all of the different model providers and the different models and how it works with different models. So
Erica (12:42) Exactly. And then the next problem you have with skills, me and my colleague Rick uh so his callbreak is a wonderful friend and colleague, but he has the next problem, right?
That he's been actually thinking about a lot and he been talking to me about it, which is okay, well say I can get hold of these skills from these vendors and whatnot.
But if I come if I have loaded multiple skills into my agent, what if they reduce the quality of the performance because together they make it worse, right?
Because skills are funny like that.
It doesn't mean that okay more skills is better sometimes how do you pick the right set of skills for the job to be done and uh
I uh I need to catch up with Rick again but he said he's been doing a massive like skill evaluation framework that's to try and figure out um just like from our perspective just even like internally where I work like how can we have skill combinations that we know are powerful that isn't causing worse outputs.
So he's like running through like an eval of like which combinations, and the combinations become so many, and it's a headache.
And as a vendor, if you provide skills outside to say like say AWS provided AWS skills, right, you can't properly, as you said, first of all, the model, the agent that the user is using.
Am I using codex, am I using Claude, whatever?
And then on top of that, they don't know what other skills I've loaded in to my agent.
And what if it's gonna mess things up?
Demetrios (14:27) Yeah, there's like these hidden skills that you may have already on your side, but then there's also you're at the mercy of whatever provider to create the most optimized skill for that situation.
So some folks, like James, I'm sure at AWS is day in, day out trying to optimize those skills and trying to optimize the MCP server and trying to optimize the agent experience.
But other folks and companies might not actually care that much.
And so if you're just subscribing to the skills that they're using, then you kind of have to pray that it's A gonna work and B gonna work efficiently and in a way that doesn't blow up your cost.
Erica (15:19) Yes.Oh, it's okay.
I'm gonna do like a fun little segue here, because this might actually be quite a lot.
Demetrios (15:27) It's cool like a segue into Jev. Are you gonna throw in a Jev?
Erica (15:31) Yes, because we know we're talking about decision making and how we can make a decision, what is the right thing to do, and blah blah blah blah blah.
Okay, I had literally not had this particular thought before we start talking about this.
However, what if you used in your agent like hardness system however you want to call it that
as you load in skills what if you can use Jev here's like the task here's like the prompt here's the context as we call it in Jev here's the context about which we're reasoning what if we can say like okay we can
grade skills like we can have we we set our labels we want to say like useful help like is this helpful we can have like a grading scale maybe one to five I don't know I have not thought this through what if we can use Jav to help us load the right skills in
like make the decision about what skills are we loading in for this task a job whatever that could be quite interesting
Demetrios (16:37) Yeah it's almost like a little bit better tool search or maybe it's not even better it's just a different way of doing tool search because I have seen that
There is a pretty common pattern. Like how is tool search done already?
Well, I think a lot of folks were putting classifiers in and it's funny, Jeff to the whole ML world.
He's like, Yeah, of course.
We've been using classifiers since twenty twelve, or Oh boy.
Erica (17:09) That was my reactions when I looked at Jev.
I was like, Okay.
But first of all, this ain't a new concept, right?
I felt like we are however.
Then I was like, okay, but why are people so excited?
People I think some people are excited for the wrong reasons, especially people who go enough when it's like it cannot hallucinate.
It can also hallucinate in a different way because the confidence score of its great like classification can it can say it's super confident, but actually not maybe it isn't that great confidence score.
However,
Demetrios (17:45) Where it's confident for the wrong reasons, like you don't understand where that confidence comes from.
Erica (17:50) And we disagree with that score, as a human reviewer, maybe, right?
So it's like this you can still get wrong things.
It's not like this that is a misconception that I see out there.
But yeah, I had this reaction.
It's like okay, but this doesn't feel new.
And it however, what I realized that is so beautiful.
And it's a amazing lesson for everyone building any platform ish solution like that other people are adopting.
the developer experience of the interface to make this request, I believe it what is what makes it exciting for people.
They can reason about classification, they can easily access this power.
So you know, you said like, oh, we have garden classification in machine learning for a very long time.
Well, that was really hard.
The work to get to that classification system and then be able to use it and incorporating it into your system was hard.
And I think the thing people aren't actually praising Jeb for is ease of use.
It's developer experience and I think that's what makes it exciting.
Just because now you can hit an API and get a classifier.
Is that your idea or it's like the API and it's a the way the API is structured.
So it's like hey, this is like my context, here's the question, here are you don't have to even yeah, worry about infrastructure setting up the whole system around it.
You don't have to train a specialized classification system like it's L M native way of doing classifiers, and we are now used to working with prompts and we're used to working with the context windows.
So being able to translate that into the classifier is what the unlock is.
yeah I I think it's just also like how you structure it like here's the question here's the type of answers you can give however I think I have not looked into this in detail but
I suspect that there are and there is an art to using Jeff like how do you structure all of this well so that you actually get value out of it and
that that's a whole other topic, and I don't know the answers to it, but I think it's will be an exciting thing to see emerge sort of best practices.
And so anyway, that that's quite fascinating.
What are some of your favorite use cases you've seen around with Jeff?
Okay, I they have like the standard stuff, which is like classification, but I think what could be really exciting.
So have you ever chatted with an AI like a customer assistant, and you get frustrated and you're like, I'm gonna speak to a human.
Yeah, please, human, thank you.
Human, all caps.
No, all caps.
Yeah.
So very annoying, right?
And I think this is actually one of the use cases where Jeff is quite interesting in so say you have your AI customer assistant, And then Jeff on each turn of this customer assistant call or text, right?
It on each call, I mean not a phone call, I mean a each turn, right?
Yeah, on each turn.
Can evaluate whether a human should be looped in.
And trying to loop a human in before you get frustrated and write give me a human now.
So it's like So they're asking the same thing for the third time, I should probably take in a human.
Yeah, or this is literally something that their AI customer system cannot do.
So I need to loop in a human.
So instead of waiting for the AI agent to sort of fail, you can proactively loop the human in and you can give people a better customer experience.
So I'm excited about that sort of use case where AI can effectively become less frustrating.
Yeah.
Well, speaking of which, there's a lot of voice agent use cases that I've been seeing.
One was just asking your computer to put different graphics onto the screen when you're video recording yourself.
So it's like create glasses and create a hat on the top of my head, and you could say that and it would put on these different graphics.
And that's cool.
But what I think in the same vein as what you're saying is really interesting is the piece where if you're running a voice agent and you need to potentially have information
that the human is going to ask for, or you need to go and fetch different things a few steps ahead of the human because it takes a little longer to do that.
then Jeff can classify if it should go and like send that worker out to get that information so that you have it by the time the human asked for it, because you can kind of tell, oh, this conversation is going in this direction.
And Jeff can say with X percent confidence, it's going to go down this path next.
Let's grab all the relevant information so that the voice agent has it when it needs it.
Yeah.
Okay.
I think I haven't thought about that use case.
What I had thought about about voice agents, because I'm quite passionate about children and children's learning.
And I realized that my seven year old niece, who honestly, she is not a big fan of maths.
Okay, but I wanted to make her enjoy maths.
And I thought, what if I can create a little tutor for her, like a little okay, I called it like a mochi cat because I think the Japanese style now back to Japan of teaching maths is really good.
And I so I wanted to do this, but I ran into this problem of I wanted to make sure that A, she couldn't ask other random stuff and get weird like guardrails.
Yeah, so I was like looking at guardrails and I was looking at things.
So this is like weeks and weeks ago.
And it just feels slow.
So what frustrated me was that when I wanted to make sure I sort of classified, if you will, I wanted to understand how is this about maths, actually, I don't even feel like guardrails is needed because my main thing is this are we still on the topic of maths or not?
Right?
And I tried a few different ways and it always felt too slow and I know her, she would get frustrated if she have to sit and wait for the response to come back, right?
She wants more instant gratification in her interaction.
So I think that Jev, with its fast classification, could be quite good.
Like I have a very like this question from child, we get that from voice to text, right?
And we can could we make a very quick assessment with a confidence level of is this question about maths.
Then on the other side, we need to make sure we are responding appropriately for a seven year old child, right?
And are we still on the topic of maths.
So that I think at least can I I thought as someone like me, that becomes such a valuable like service and way of doing it.
Yes, it doesn't replace the entire idea of guardrails and all of that, but I think it can help, especially with these smaller solutions.
Why am I thinking about this?
Because my first reaction was, by the way, with Jeff is who needs to make decisions at this speed?
Okay?
Because people are excited about like the structured responses.
And I'm like, dude, you've had structured responses from Anthropic and OpenAI for ages.
So ages.
Don't be excited about that.
Because that isn't what's exciting.
Like we've been able to do what Jeff does already with the existing inference, but then that is slow, right?
It is genuinely quite slow.
And slow is fine for many because it doesn't matter so much.
And I was like, oh, we need fast at scale, like think big enterprise scale things.
But then I was like, okay, but who what small projects is it good for?
And that's how I arrived at my maths problem.
How can I have a small thing that needs to set spade.
Uh money.
It reminds me of back in the day in 2022 or 2023, right?
When we were starting with LLMs and trying to figure out how to set up the system around LLMs and make it back in those days it hallucinated so much more.
And it was really it felt like the raw LLM that you were getting there was no nothing else in the harness, quote unquote.
Now it feels like it's much more of a system behind that API when you call it than uh just like Just like yeah.
I guess you can still get just the LLM uh if you don't set you want it almost like quote unquote naked.
But even with just hitting the LLM, they do a lot of stuff behind that API that we don't see.
And in the early days, they didn't.
And so a lot of folks right from the get go would throw a Burt model in front of the LLM call because they wanted to catch if a user was trying to prompt inject or trying to do something to make it say something that shouldn't write from the get go.
And it's a lot cheaper to catch it before you make the LLM call than after when the LLM has already given you the response and you're like, oh whoops.
I didn't want you to actually respond to a that type of question.
Yeah.
So yeah, JEV could potentially help in those topics.
Jesus That all over again.
It is like bringing that design to the masses.
And I've seen folks that are using JEV as computer use.
Uh like augmentation, I guess is what you would call it.
Because there's only a select amount of tools or things you can do on a website.
And so Jev is able to help speed up what you are able the action that the computer use takes Yeah.
the website.
Like to make the decision about which one to take.
And where wanna go what yeah exactly that was something I'll be thinking about as well not necessarily computer use, but just like a tool selection.
Wait.
Computer use is one where you you see that oh cool.
It's able to just choose the next step out of a probability of things that I want to press.
Do I press the search button?
Do I type in something into the search box first?
So it's able to do that really well.
I this kind of leads me into some of the awesome conversations I was having in Amsterdam with folks when we did the agent con and MCP con EU.
And I don't know if you saw, but I was live streaming that.
I know you and Live Street.
Yeah.
It was very cool.
And we pre-recorded some podcasts on a tandem bike, so I managed to get some of my friends to ride a tandem bike with it.
And we strapped a bunch of GoPros on it and then we recorded a conversation.
Yeah, it was it was kind of fun, but I almost ran into stony cars and so many other bikers.
Maybe you should have gone to like not be in the traffic.
Okay.
Okay, learning.
If you ever do this again.
Like so that information would have been nice to know before we started at Central Station on the tandem.
Like, oh, this won't be that hard.
And then there's mopeds, and there's other there's a million bikes everywhere.
Anybody that's been to A Sam knows that bikes rule the roads.
There's pedestrians.
Better high.
There's all kinds, and then there's big trucks, and there's people that are just stopped in the middle of the road with their cars and so you have to maneuver around them and so it it was a little stressful.
Uh I had I think it's stressful even if you're not like recording anything like Amsterdam If you're not even.
on a tandem bike it is stressful.
Walking is stressful because there's so many bikes.
Exactly.
So that was that but one of the cool conversations that has stuck with me since that event that we did was all the talk about Web MCP.
And I'm really excited about Web MCP because of exactly what you were saying.
Can we give all of these websites these tools to expose to computer or browser agents?
And you and I have both talked at length about how stuck we are about computer use and browser agents.
Yeah.
And so this feels like web MCP feels like a natural way to start exposing more capabilities to the agents.
And some of the learnings that I had from these conversations with folks about how to build web MCP servers for your website, it was very much like think about the top four user journeys that people take when they are on your website.
When a human lands on your website, what do they generally do?
and then expose those journeys as tools not necessarily like a million different tools that oh you can click you know on this product or you can click on that you you almost have like
these different happy paths that you can take and make those the primary tools that you put in your web mcp server and I just feel like JEV plus web mcp is a killer combination that I want to explore more and see how that works, and if it is super effective and super fast,
that might be actually really cool.
Yeah, it's quite interesting.
Yeah.
And also okay.
The thing with Web MCP and like MCP in general, that is like on my mind is I mean, we've spoken about this before, but I feel like we're playing catch up in authorization and access control and
I have thought about it and this actually's the funny part for me is I have connected the dots in my head about why JEV is so exciting for people because it's usable, right?
It's the interface is good.
And we are we have a problem with authorization for MCP because of OAUF, okay?
And because OAUF is not designed for this, but it is the most adopted authorization system, right?
Like approach to authorization and adoption is so critical.
So I think Yeah, it's almost like this going on a circle.
Trying to rock in a hard place.
Because it's the most well adopted, but it's not enough.
Even the one thing that I heard people saying at Agent Con was that the guy who wrote the OAuth spec, he has come out and publicly said that OAuth is not enough for this new agent paradigm.
Yeah, and like OAuth has been struggling even before agents with so much and it because it was just talk about like one moment.
What was OAuth created for?
Before OAuth, before consent screens, for those who remember the internet before consent screens, right?
How did you connect things like applications?
You literally gave your password to so if Facebook needed to integrate with whatever other system you had, you literally had to give your password.
So Outh was designed to stop people giving their passwords away to third parties.
Like that was what we were solving solving this thing for humans, for humans to stop giving passwords away.
And it has been evolving and evolving and evolving to fit the API era and API distribution era overall.
I mean that kind of fitted the original use case fairly well.
But even the world of the explosion of API distribution and the way we think about multi-layered systems has always been a challenge.
And now we introduce like agents which are deterministic software say that is tricky, but but I do think it's quite fascinating because authorization is only gonna work if both of us agree to operate within the same framework of like how do we establish authorization.
So we have to both be willing to agree this is how we do it, and oh there's been there.
And uh people are doing a lot of work to evolve it, but yeah now I was thinking about Web MCP and I'm like dude like
it's like it's exciting but it's yet another surface area of how can we audit authorise and manage access
and it just hurt my head yeah the thing that I will say in Agent Con is that almost everyone of the booths that I talked to was saying that exact same thing where it was like, hey, we need to get off, locked down, because with MCP, they're making the steps in the right direction.
They support OAuth and all of that good stuff.
But it's not like it just comes out of the box with the standard, and you as the person who is either creating the MCP server or consuming it, you have to be cognizant of what you're getting yourself into.
Yeah.
And it's it's so anyway, so fascinating.
And then like on that same topic of security.
So you know we talk we've talked before about how everyone wants an agent that is in training going and committing crimes on the internet, right?
Like every lab wants to have a naughty agent.
Yeah, it raises the valuation of their company.
it's great.
Uh got Naughty Agents.
Then because it great makes great news stories.
And then I thought about it a bit and I think it relates to this off part about MCP as well.
I'm gonna make a weird analogy, so you're gonna have to bear with me, but you know how it goes.
So imagine you have a bunny rabbit.
You're having it bunny rabbit.
We're already starting on a great foot bunny rabbit going.
Yeah.
And uh the cute bunny rabbits they have wonderful like little cages if you are one that puts your bunny rabbit in a cage.
So for example, when I was little, I had two bunny rabbits and they lived outside in a beautiful like little house my dad built with a beautiful uh big outdoor like cage, if you will.
It was using all of the c you know those um they're quite You use them normally for like hens and stuff, I put it.
It's like in metal method wire.
Almost.
Yeah.
Yes, it's metal in the wire.
Yeah.
Okay.
Great.
This works very well for the bunny rabbits, by the way.
We had to put it at the bottom as well, because the bunny rabbits can dig.
So we could it we had an a situation where the bunny rabbit broke out like these naughty agents, right?
Because we hadn't thought about the part the bunny rabbits can dig.
They can.
So then we had to put the netting underground too.
So no more digging.
Only supervised digging allowed, no unsupervised digging happening.
Okay.
Cool.
So now we have a great cage for a bunny rabbit.
Bunny rabbit stays home, doesn't get we live next to the woods, does not get eaten by predators, a mason.
Bunny Rabbit stays safe.
Okay.
Well, we're very happy with our cage.
But now we got a bear instead of a bunny rabbit.
But do you know what we do?
We just got a quit?
No, I did not in real life, but just like imagine you got a bear.
Oh I did not have a bat.
Well, Sweden.
I did not realise you were that hardcore.
That would be amazing, right?
We got a brown bat as a pet.
No, the brown bats were wild in the woods.
But you know, but imagine we're not.
We still have chicken wire.
It doesn't work.
It doesn't work.
Now it it's a really useless, Kate, even though we were really good at putting the netting at the ground as well.
So I think for me Yeah.
So the the bear is strong enough.
And maybe even if uh we had a dog though, maybe even just we imagine put the dog in there.
I don't know though if the dog would have gone out.
I think she's too timid, or was too timid, she's no longer alive, because that would be a crazy long dog life.
But anyway, imagine then you put a much stronger, like a bear, uh animal in this chicken wire cage.
They're gonna just break out.
And I feel that when we talk about whether it's off for agents accessing MCP servers, or we're talking about how do we create safe sort of sandboxes or environments for the agents.
Bringing a barrier out.
I like it.
I feel the problem is not the agent, it's we are shit at building cages and the safety mechanisms.
We have not upgraded how we approach the cage.
Because we're trying to use Off, which in this example is the chicken wire, with a bear which in this example.
Um but be something like so you've been able to land the plane on that one, but I like it.
Now I see where you were going.
I mean apart from that I cannot get a bear to be my pet, but that's a different problem.
They have bears in zoos, though, and they do not use chicken wire for those bears.
They tend to do all kinds of stuff.
Exactly.
If we extrapolate that out with our agent world, it is possible to keep the agents in captivity and upgrade our OAuth settings, or maybe Maybe yeah.
not defense at all.
Yeah.
And I feel like there's you know, there's two analogies here, but I think it applies to a lot of world safety, whether it's sandboxes or our oath mechanisms.
It requires some I guess as you say, like the OAuth dude said, right, we need to think slightly differently, whether it's and if the same comes down to infrastructure and how we secure these environments.
We've optimised for human computer use mostly and automated processes etc we yeah I I think we when people are afraid that AR agents are going rogue I feel well I'm afraid that the humans are not spending time thinking about how to build a better cage.
I like that here's a question that I had come up and it was posed to me when I was in Amsterdam, but I would love to get your take on it because I've been thinking about it quite a bit.
It was Do we feel that computer use is just a stopgap until everything turns headless, and we only use MCP and APIs for everything through our chat terminal or agent chat, whatever you want to call it.
Like the gateway to the internet or to the machines.
Or is it going to be that we are forever going to have computer use as a valid way of doing things?
And it could be that we're talking in terms of 10 months or 10 years.
Ten years down the line is every single website or every single service going to be turned into just an MCP server so that agents can consume it.
or are we still going to have folks that hold out and I'm thinking about the German government.
I don't see them ever having an MCP server for my driver's Okay.
appointment.
Yeah, so I'm always going to need computer use in that case.
And then I was thinking about it that well if computer use is good enough, why it doesn't give us that push to need to transfer to MCP servers.
We don't need to have everything go headless because if the agents can do what they want to do within the current paradigm, then what is that forcing function for us to go and actually create that world of everything is headless?
I okay.
What is your thought?
Because I have thoughts, of course, because I create thoughts faster than my brain.
I mean, I think out loud, but what are your thoughts?
I think I'm going I'm leaning more towards computer use is so good that it's going to make it so we do not need to go fully headless.
Because going fully headless would be working on something that isn't necessary.
Yeah.
Okay, I do not think I agree with that.
I do not believe we're gonna go fully headless.
And I'm gonna ground that in the fact that despite API surfaces for many things, we did not end up headless.
Even though just keep in mind that I think it's at least I think it's more than three quarters of all traffic over the internet is API traffic.
Okay.
But the part that I'm thinking about why I don't think we're gonna end up in headless, and what really defines headless.
So first one is I think we may be getting better user interfaces because I believe the computer use and humans can end up in a cobot situation.
So do you know what a cobot is in like a factory or a warehouse that's a robot that works alongside a human.
So the human doesn't have to do like, for example, cobots have been used especially in manufacturing and warehouses to reduce repetitive heavy tasks for humans that actually damage humans' bodies, stuff the humans were not built for.
So we build machines to work alongside humans, not instead of, but alongside humans.
So I do think f we have a workaround with computer use with things they should be an API, but isn't we can work around that.
I love it.
LinkedIn anything that we do on LinkedIn, oh my God, why can we not grab that API, but And dude, like even Gmail, I was angry at Gma yesterday and I had to resort to computer use for beautifully formatted emails because I don't know why, but
making beautifully formatted emails either they don't make it easier, you have to work around it.
I think they can do a better job.
Google, please make better Gmail interactions overall, so I can make beautiful emails.
Victor's just like I don't want to do the whole HDML CSS stuff after body, like I do just wanna have like bolding and structuring and beautiful stuff without just text formatting beautiful.
I don't think that flashy beautiful.
Yeah, and that was a problem.
And uh anyway, I think though that we can get to a point where maybe user interfaces and computer use is more for when we want to work together with AI versus AI working instead of us because when AI is working completely independently of us, it seems a little bit wasteful to
rely on computer use in the way it is.
Then I have this other thought on how we can bring experiences to people.
So you know MCP apps which have not taken off the way I would like them to because actually but the problem is actually like cloud and open AI and all of that jazz because they aren't like the the way they don't default to them.
It's not like publish one.
You can't like you have to be very special to get to even publish one at the moment, which bothers me, but different I think that some websites and like interfaces are simple and small enough that they should be able to just be an MCP out of we we can present a small bit of user interface in chat
that that's the other partn the other funny thing I thought when you talked about humans having the main interaction point
chat like dude there are humans are really bad at explaining what they want like I know that there are a lot of humans who've gotten really good at using language to describe outcomes thanks to the rise of like AI because they have to use language to express it, otherwise they will not get the outcome.
But because humans are pretty bad at it, I think that the notion of interacting with websites is gonna remain simply still valid.
Because we also need humans just to do their thing because they're bad at explaining.
The other fun thing is that you can obviously record like your own actions like you can record you can use like Claude and like record what you're doing so instead of having to use words to explain you can show and if you make it headless a human cannot show Claude right so
how it's always going to be a need for that oh I didn't think about it see I'm thinking like how how is this humans need to be able to show
just I I may I think we may get to a reduction and maybe we even get like companies.
You know how you said earlier with Web MCP that what are your main like use of flows?
What are your things exposed then?
And when you talked about that, I thought about all of the things that people do when it comes to building apps or websites.
I think this idea of really focusing on the core flows.
It's sure, like it's so easy to build and build and build that you forgot what the core flows are.
So I think what's kind of interesting with this, like, okay, well, where do you start with web MCP?
Maybe start with these like high value core flows that people do.
And that might just be a valuable exercise for people in general to think about, regardless if they're building a web MCP uh solution or not.
But maybe we will get leaner and cleaner like user experiences as a as maybe the more like corner cases and things move into headless because most people like we start catering to the eighty percent rather than weird use cases.
I think you are touching on something that I hadn't thought about too, which is how nice it is when you stumble upon a beautifully designed website.
where you get that copyright that captures you and it's with the design that when you scroll you go like wow they put some thought into this one yeah if everything is just through the chat window you're not going to be able to experience that as you would through a website
even if it is with MCP apps MCP apps right now doesn't give you that full it's not like it takes over the whole chat window
to recreate that website in your chat.
An experience, yeah.
Sure, we do have the browser on the right side of the chat where you can be co working with the browser there, but it is and maybe that's what it comes to is where all right I still have the different websites and I've got the chat as my first player.
But you know, I'm kind of looking out and enjoying the scenic view as the websites are coming up and as my agent is doing things.
Yeah.
Oh that's got like okay, I am guilty of actually enjoying watching like computer use take place.
By the way, like I am guilty of enjoying it.
Uh I think it is similar to my like Tesla self-driving experience.
Like I'm just excited to see what the computer is doing, which is entertainment in a weird way.
But I I have to admit that I do enjoy and then I and then when I catch myself I'm like, am I just wasting time?
Like yesterday when I wanted to have it make a beautifully formatted email, right?
Could I make a beautiful email myself?
Yes, but I just wanted to like it's so much more satisfying.
When I know that does it and you watch it.
It's just like exciting.
It's like the w uh maybe the wonder will disappear.
Maybe this excitement of trying different things and see how it would do it, what it does will disappear with time.
so far I well, I'm gonna enjoy it for as long as I enjoy it, and we'll see when I get bored.
But on this one though, I maybe can wrap up if I can tell you about the fun part about design because I love hiding Easter eggs in design.
Uh Tell me.
Yeah.
Okay.
So I got to have the wonderful privilege of designing a website recently, like the Agent Router website.
And because we had to like rebrand it if we're joining the AIF and stuff like that.
But the most exciting part, what do you get to do when you design you get to hide fun Easter eggs?
What'd you find in there?
Okay.
Maybe don't tell us we'll have everyone go and check out the website.
I th I think like you have.
I am not sure where they were gonna get it because it's so yeah.
Unless there are millennial women.
I did give a beer in cages surrounded by chicken wire.
No, it's not.
I I'm a big fan of Taylor Swift, and for those who don't know Taylor Swift is releasing a new single.
But whatever.
Uh Taylor I'm a big Taylor Swift fan.
So I decided to make the entire colour scheme of the open source project Life of a Showgirl.
Oh no way.
You should copy that.
So it's like Life of a Showgirl famed.
For those who care.
I did tell people, some other people ahead of time, that I want to do this.
And they're like, I like the colours.
Okay, but just so you know, it's Taylor Swift, Laugh of a Showgirl.
And they were like, Yeah, I don't care where you got the colours from, Erica.
Like and anyway, I think that that's the fun part about design, like and why it's so fun when humans create things over all is that there are these weird backstories, like these weird ideas people have.
Like, I think that's often what's exciting when you see a design that isn't AI generated completely.
I will say one of the my favorite things to watch on TikTok is folks that are going through some of their favorite designed websites, and they showcase these really high quality websites that a whole design firm spent months slaving over and creating and bringing into this world.
And yeah it I think about that as almost like the twenty twenty six version of that old website we used to have Stumble Upon.
Do you remember that where you could just click a button and it would be a little bit of Oh yeah it's like I'm lucky.
Lucky Yeah.
that was the Google version where it's like feeling lucky and it would just take you to the first link, but Stumble Upon was even I think
like right around when Google was first starting, and it was a website that you would go to, I think it was stumbleapon.com, and then you would just click on a button and it would randomly take you to a random website that is out on the internet.
Okay, nice just run okay, that's just rolling the dice.
safer.
Yeah, because I think they curated a whole lot of domains where they would put them into the the random number generator.
And so when you clicked on it, it would take you to a cool website.
And it was always interesting websites where it would expand your mind on what is possible with the internet.
Yeah, because remember, this was probably like 2003, 2004 that we were playing with this stuff.
Maybe 2005, I can't remember exactly when, but it was early enough to where it was like, Whoa, there's a game on the internet.
Oh, cool.
Or let's, you know, let's try and play it.
Or whoa, you can move these pixels around.
And so it was able to open doors into what the possible was.
Do you remember this?
Okay.
I don't know we have to wrap up, but like, do you remember this thing where was oh I could not remember what it's called.
It was a way for people who were really good at CSS to like show off the CSS skills.
It was like you worked with the same HTML and you only applied uh CSS to it, and you could see this whole collection of it's the same content Uh of like Dojo or something.
I do remember that.
Yeah.
Remember this, and you could like and that I think was quite interesting, where people could like make two wildly or you know different experiences not by changing the HTML or by changing the CSS.
Just the style.
and I think that is really wonderful in like how you I it's like CSS artists almost right.
They're like such creative.
And it's how you express your idea that you're trying to get across.
And inside of the chat, we don't really have that expression yet.
And so I agree with you that we're not going headless anytime soon because we will completely lose that whole medium and I don't think we're ready to let that go yet.
And it is a much richer way to express ideas.
Yeah.
I think this interesting like what an interesting question.
What will our experience our digital experiences become?
Right?
Yeah.
Dealing with just the chat feels very boring.
Just Mars down files.
Like really?
That's the best we've got.
We've gone backwards and now we're just going back to 1996 when we're looking at plain text files.
That's it.
And what and if we come back to like the internet, right, and the ability to connect with people from like different parts of the world, just the fact that we sit on two different continents recording this is, you know, the excitement really of the internet.
Yeah.
And
I think about
what how is that going to forget about getting computers to do work for us but how is oh
I think something has to evolve in the social media space by the way I feel it's social we has not gotten enough love recently but in terms of improving our human to human interaction over the internet because I feel like all I got is ads at the moment and I'm exhausted by Facebook.
I don't go on there anymore because I don't like the experience.
Not because I'm against Facebook, by the way, it's just like I get exhausted just looking at it.
Say it's bloatware.
It's Microsoft Excel.
You know, it allows us to be a opposite of a beautiful Yeah, the five years.
side.
We it's all of a sudden it's become the platform that you have and it happens to the best of us, but it's that inertia of all the different things that the platform then adds on, and if you're not ruthless about what you take away, then you end up with something like Microsoft XO or Facebook.
Yeah, feature gluttony.
Payful just exactly feature gluttony, feature float.
So well, next time that we chat, I want to talk about this other idea that I had been hearing at both AgentCon and MC Con, and it's the idea of not really understanding the traffic that is hitting your MCP servers and not knowing if is it the same agent?
Is it the same person behind that agent?
Is it two different agents, but it's working on behalf of the same person?
And we do.
what did that agent really want at the end of the day?
Did it get what it wanted?
So that whole idea of like understanding the traffic when your agent is trying to consume something from your like if you're an MCP server provider, like how do you even Yeah.
There's a lot of talk around that just because you don't get to know what the prompt was.
You just know that the agent called a tool from your MCP server.
And so you know if by calling that tool that agent got what it needed?
I think this is I think this is exciting because it also it's related to the authorization story, which I think.
Alright.
Well we'll cap it.
We'll save it for next time.
Always a pleasure chatting with you.
Thank you for being back.
Adios.
+ Read More
Watch More
Illogical Logic: Why Agents Are Stupid & What We Can Do About It // Dan Jeffries // Agents in Production
Posted Nov 15, 2024 | Views 1.5K
# Logical Agents
# Kentauros AI
# Agents in Production
Why we built PydanticAI, and why you might care // Samuel Colvin // Agent Hour #2
Posted Dec 19, 2024 | Views 4K
# Pydantic
# Agents
# Agent Hour
# AI agents in production
All About Evaluating LLM Applications
Posted Sep 28, 2023 | Views 903
# Evaluation
# LLM Applications
# Exploding Gradients