My Secret Agents (IYKYN) | Vijay Krishnavanshi | ETHMumbai 2026

Hey everyone. All right. So, let's go to the slides. Okay. So, today uh first I'm going to share like what I'm going to talk about uh uh here.

Uh I'm Vijay. I'm one of the co-founders of FileW. And uh today I'm going to talk about the evolution of the internet that has taken place in the past couple of months. And it is also going at a break neck pace and uh things have evolved much more than what we anticipated. Like you can see the first line says that humans are no longer using the apps that you guys have built, right?

And it's also there is a different kind of person that is using the app that's not you and me. It's someone else. You might have guessed by the name. It's agents. And uh what they can do is uh very wide in terms of the range of things that they can do.

They can read different emails of yours. They can uh take actions on the behalf of you. They can do uh things that uh like you would take a lot longer to actually come up with to do that in a go and uh like one-shot it, right? So this assumption that we had that there will be a human on both ends of the internet is void. It's invalid.

Now so what do we do next? Um like before I go and explain that I would like to tell a bit about uh Fileverse and uh what we do and uh like why do we get excited to come to work right so until 2010 there was this thing which is like onethird of the internet was peer-to-peer but that has changed and today us all included and uh we trust this three companies to handle, manage, save and uh allow us to share our data with others. What are the things that are included in the data? You can uh think from the like emails to docs to sheets to files. Everything that you have as a digital footprint is something that these companies store, share and allow access to to you.

These platforms put at risks people's creativity and uh also their privacy and also their digital freedom. If you share something and that that doesn't agree with the rules that are created by these world garden companies, you can be kicked off from those things. And uh what are we doing? like we are a group of small uh like we're a very small team and uh everyone has a different uh X factor in the team and what are we doing is we are building an alternate to that and uh we're building open protocols using public networks versus the old centralized internet that uh we all see and use but today it's not about files it's about the change or a critical change that's happening on the web and I'm going to uh talk about those in the next slide. One second.

Yes. So, internet has a new kind of user and it's not you and me, it's agents. The web when initially was thought about it was thought about for people and bots were treated as spams and that is why you have captures uh everywhere on any website that you see that you have to prove that you're a human by clicking a button right which uh is insane but that baseline has changed and uh agents are now first class users and they can do things on behalf of you they can make API calls they can read documents of yours uh they can send email to someone on a button click. They're not just a feature. They're a new class of users and most app including yours are not designed for them.

I'm going to highlight in the talk that I have uh the different principles that we came like thought about when we were building our feature uh like API access uh in the privacy first product called DOC.NE. Uh next slide. Yes. Now why is it different from the previous automations that have happened on the web, right?

Uh meaning uh you have seen different frameworks that have come up for bots. You have seen different frameworks that have come up for uh UI development and all of these things uh o and uh everything included. But why is it different is they have gotten this new capability uh like basically a code that you write on the web has gotten a new capability to be autonomous and make decisions that was not possible in the bots. Earlier when you were writing bots, you had to be rule-based and uh like you just had to rely on the compute power, right? But now they can take decisions such as testing your app, making sure that whatever they are doing is working and everything else.

Now, but it also comes with the risks, right? Meaning if something can run a thousand or like a million times faster than you, they can do a lot of damage if they're compromised. And uh these agents they don't need break they don't slow down they can ingest your entire account and uh create a summary out of it that's a boon but also a threat right so I'm going to uh share in the next slide the problem with how we build today uh there is a problem with how we build today the current default that each of us follow right now is you have an app you want to add an agent support to that app what you do is like you share the entire accounts access uh take uh like you can take any company that you have in mind you can take Figma for example all of these things what they do is like they enable oneshot access to your agent and entire account is now controlled by that bot right or that agent but this is wrong but the and we all like uh whoever is a developer knows this by heart that this is not the right pattern but this hasn't been like explained very well or like you know defined very well as a problem. So and uh this is a note for anybody that is building tomorrow is don't share the entire access of like don't expect your users to share the entire access of their accounts. Ask them to only share what is the relevant data.

I'll share the principles in the next slide but uh and this is the change that we need to think about when we are thinking about building with agents. Now there are these are the four patterns that we were able to notice in all the different kind of uh use cases that were there. So basically first you need to make sure that you give selective access since it can uh make sense of any amount of data you make sure that you give it targeted access. That's the first thing. Second is bake in the kill switch on the admin level.

Meaning whenever the agent goes rogue or you don't agree with the action of the agent, you should be able to press a button and stop that agent from working any further. You should have verifiable logs meaning what do you whatever the agent actions are, they are recorded and they cannot be tampered with by the agent itself. Then you also need to have versioning. Why the versioning is because let's say there is a damage that has been done by an agent. you should be able to roll back to the previous state and that is why this is an important parameter to keep in mind and it should be a full account state versioning.

Now let me show you that uh how we use these principles in building uh our APIs and it's not just theoretical right uh next slide please. Yes. So what we built so basically we created the sealed space inside the file account and this sealed space comes with uh like the scoped API key access which can only read and write data from here and it has full auditable locks and can be revoked at any time. So all the four principles as you can see are applied in this uh next slide please. So now this agent never touches your full account.

Okay, why that is crucial is because as a human you do changes and you do things but you want to make sure that the agents can only see whatever it needs to right but not more and this was also highlighted in the keynote talk by Vitalik as well uh earlier today uh and the stack that we use we use safe 4337 smart contracts IPFS and UKAST you can read about this later on uh I'll go to the next slide Now yes now while we were building this there were some challenges that we had faced right uh and uh these were not uh trivial challenges these were challenges like how do you keep end to end encryption and uh make sure that the API gets access uh to the data that it needs. What we did is like we created an npm package that was able to do encryption on the API boundary level so that as soon as the data gets the API gets the data it gets it into the encrypted format and then everything gets stored on chain and uh pushed to the file ecosystem in general. And then this there was another challenge that we faced and this was uh something that was funny and we found out later on uh is uh when we built the npm package we had to also think about different runtimes where it can run and uh uh make sure that it is uh runtime agnostic and with different type of crypto APIs that different platforms provide it's not easy. Uh next slide. Yes.

So another important thing is like with end to end encryption if you have API keys um you need to be able to rotate these API keys and uh think about the challenge of like needing to rotate the API keys for an end toend uh encrypted uh ecosystem without losing access to the previous data that you have shared with the uh agent right now. Oops. Uh can you go back? Yes. Another important thing is from the UX level which is to be able to let to reduce the threshold from a technical user to a normal user.

We had to work on one-time deploys to Cloudflare and Heroku and all of these things were thought from the first like from the user's perspective. Right? Then uh one of the biggest challenge one of one of the like you can say side quest that we had to also think about when we were building this is like you're doing encryption on different platforms. uh for example you are doing it on the web level you're doing it on the uh uh system level and you want both of these things to be able to work together right and uh you want that whatever you have done the changes that your agent has done is also visible on the UI and you are also able to continue doing edits on top of it if you need so that was another thing that we had solved with the yjs plus markdown compatibility uh next slide please now how am I using these the APIs that I just told you about or what are the different use cases that I have in mind. Uh it's for example like I'm using claude and uh I'm I I have opened it into a repo.

What I do is like I ask it to generate the documentation or let's say the highle overview of the repo and uh give me a do link uh which is also a skill that we have created. You can find it on clawbot or different uh platforms. You can just go to uh the npm package add the rate file/ API as well. Uh and uh what we do is like we generate the documentation and the link is shared back which can be shared with anyone that they want. Uh next slide please.

Yes. Another use case here is the telegram clawbot and dodo. Um in this what we are doing is like we we have connected clawot with the telegram and whatever messages we're sending to the clawot on the telegram it's also doing those changes on the API level. Now doing this is not that big of a task but doing this in a way where you have the sealed space and you have the complete control of uh that sealed space without leaking access to the uh accounts and stuff like that. That's the difficult things uh in these and we also try to reduce the level of technicality that the users need.

Yeah. Now the last use case that we have is letting agents collaborate. Now in this what we did is like we uh took the same MCP server installed it on two different uh uh places. one is OpenI and other is claude and asked them to uh draft a document that can that they work on together and uh as you can see they did it. Uh now let's go to the last slide since I think my time is done.

Uh next please. Yeah. So here are the link to the slides. Uh the slides go to filever slides uh not Google slides and uh this is the twitter link uh for you guys to follow if you want to know a bit about like how file works and what are the different things that we are releasing. Thank you.