Members-Only
Recent Talks & Demos are for members only
You must be an AI Tinkerers active member to view these talks and demos.
May 05, 2026
·
Seattle
Coding is Solved. Next Up: Figure Out How to Plan
Discover how to manage AI-generated plans and code reviews efficiently. This talk explores building an agent-native Google Docs for seamless editing, sharing, and real-time collaboration.
Overview
I built an Agent-Native Google Docs to review, edit, and share plans in Markdown
Video
Transcript
Generated 3 months ago
Summary
Generating a talk summary...
View full transcript
Sn the back? Good? Better? I need a thumbs up or no. Not really.
A thumbs down. I'm really close to my mouth right now. Yeah. That was a good word. Very bad.
I'm happy to 1. Alright. I am so excited to be here, you guys. I figured this was gonna be, like, 30 people. I put in my talk.
I got permission from my wife to leave her 2 kids at home. The reason I am here is because something this little idea has been, like, worming its way into my brain and it's been beating itself out of me. And I really need input from all of you about, like, what it is and what it's good for because I feel super convicted about, like, something like this being the future. So let me tell you about it. So show of hands because I love engagement from everyone.
How many of us have 2 or more agents that are, like, accomplishing real useful things and always a long way? This could be like a claw or something and a hand shut off. Can you keep your hands up? Keep your hands up. Okay.
Great. That's a considerable number of us. How many of us have our agents talk to each other? Okay. How do they do that?
Go ahead and shout it out. Thromify. Say it again? I'll talk you'll see my demo. Okay.
Okay. Okay. Notify. Notify. Right.
Like, your message sending between each other. Maybe they're writing markdown files. Right? How many of us talk to our agents? Okay.
How do you do that? Because you may have a huge chain of stuff that's going on. How do you insert yourself into that chain in a critical in this critical time? Stop the agent, write the prompt, and then repeat it. Receive.
Message. Yeah. Yes. Yes. This chat, right, usually, if we're lucky enough to catch the agent.
So, you know, you start thinking, like, right now, all our agents are on our local machine or maybe, you know, they're in the cloud, but, like, really, they're on our local machine, and they're writing, like, the markdown files. You end up with this, like, graveyard markdown files, You're stuck in that same, like, chat window where you can talk about 1 thing at a time. And I find myself, like, copy pasting from the markdown file. Like, hey. This thing can be maybe more like that or that thing can be maybe more like this.
And Sn, you know, and I kinda find that with I'm a software engineer during the day, and I use a lot of agents, write a lot of code. I'm sure a lot of this can empathize. Like, a lot of what my job is becoming is something like an AI sandwich. And that means spending a lot of time with planning. I'm seeing a lot of edge 9, that's good.
And a lot of time in, like, the output because the the harness and AI and and LM is so good at like executing. It has a really good plan, and I just wanna see like what came out the other side. So like what's really good at that? Documents are. So I wanted human humans and agents to be able to collaborate on this, like, shared thing in a way that felt really needed to both in a way that, like, current text editors aren't.
So maybe you can get okay. Thanks. So so maybe you can get, you know, write a document all at once or something like that. But, you know, you can't do the kind of primitives that humans can do. So I I think agent laptops are kind of new primitive that we wanna, like, think about how it fits into our workflow and, like, what gets unlocked for us.
I have a lot of ideas. I have a lot more, like, kind of coming that I can't, like, I'm not ready to talk about yet, but I but I wanna hear from you. So if you are interested in this at all, I would love for you to go, like, comment a little bit differently. I'm not at least I'm not on social, so I don't know how to do that. But, like, just come talk to me in this quick pace because I really wanna understand how you can imagine using this.
Now the the code part. So, and and by the way, I'm demoing this in how much I owe, in case you didn't tell me. This is, all on Cloudflare, durable objects. My understanding is that it's fairly instantly scalable. I've been doing, like, back end scale work for a long time.
As far as I can tell, it's like 1000000 dollars for costing, like or no. Sorry. 1000000, documents would cost me, like, $10.15 dollars a month to host. So, like, it's kind of free forever, this thing, like, 1000000. And, you know, I'll put all these kind of architecture documents in the Discord for anyone who wants to take a look at it.
And and and by the way, this is all editable by agents. Right? This is just a MIRI diagram that my agent wrote. By the way, just showing, like, 1 cool feature is so, like, this text, like, 1 thing that comes up when you have a lot of agents working together is, like, who wrote that? And I built a thing that is still very early stage, but where you can tag in your agents.
And this is going to my flaw. There's a way to go through code extra or, like, code as well. And this is gonna take a long time for this running. I hope it's so it's gonna and it it may not be till the end. And then, just a a couple of things that I wanna talk about, like, what makes Asian Native work really well on the web.
I'll be really fast. Okay? 1 thing is if you wanna know if an agent is making a request on the web, and not a browser, you will get the accepts header, and it is not like the, agent. And the accept header will be something other than just HTML. Browsers will only accept HTML.
So what that allows us to do is when you copy paste this URL from comic.io to an agent, it actually sends down a full, like, JSON representation of the document and a, like, box for, like, how do I interact with this. And and, I'll stop it. It's my old name. I'm sorry. I have to answer some questions.
As I'm answering, help me break my editor. Please go to the scheduling and collaboratively edit, and I wanna get all the errors in the world. Because it's in fact right now, but I'm gonna make it better. Just a quick few questions. Do we have a show of brands in the room?
Yeah. Oh, if I can go. Just I didn't realize that. Just Alright. Yeah.
Yeah. Yeah. How do you handle race conditions between different agents, or even? Yeah. Yeah.
So it's it all comes down to how like, how Google Docs does it is, like, conflict free replicated data types. And so that's, like, individual people are, like, making these little messages that then kind of all get mushed together. And so now is I'm glad you asked that because that was literally my second thing is, like, you give the agents, the tools that they know how to use. So how do your agents edit their markdown files is with an old string and a new string. Like, replace this string with that string.
And so that's the primitive that comes up in an API. You turn that into a CRDT so it can seamlessly merge with with everything that the humans are doing, and then it all just kind of flows together. And again, you know, there's comments. There's a cool thing, like, you can lock documents and only get input. This is I can't believe this has been broken.
That is wild. This is, like, so new y'all. My product is wacky? No. I mean, it's, like, you don't understand.
I was kind of validate that it's it's real. And so, like, if I get to a point where there are a lot of users, like, I'll probably do, like, soft 2, and then I'll probably do something that, like, allows input from the community. Thank you. That's a great quote. Alright.
We're going to have, our next city come on down the shore.
Links
Tech stack
Compose Email
Sending...
Email preview
Loading recent emails...