Orchestrating attention with human-on-the-loop

Speaker 0: AI. My check. My check. Okay. AI. Speaker 0: I'm here to talk to you about notification overwhelm from all the agents that you have build staying in control. I'm gonna do a quick picture slideshow. This was me 2 months ago. I came in talking about something I build called AI, which is kinda GPT be everything you see tonight. BDD I Lead, I think this is a new AI of paradigm for how agents and people can work together. Speaker 0: It's a multiplayer markdown editor where people bring your agents run you can all jam together. This was us a week ago. We went to AIE. E2E at the vessel of the sun. E2E GPT test, please, for Seattle. Speaker 0: We're gonna make this place an awesome tech hub, you guys. And ouch. AI now, E2E think it's San Francisco and New War, and it needs to be Seattle too secure I know we have so much awesome talent here. I mean, just blocking just E2E up. How many people are GPT? Speaker 0: Yen. E2E done. E2E, at, you know, I I got a found cofounder. I quit my job. We're, you know, fully agentic, fully just Founding software as fast as we can. Speaker 0: We burn a lot of tokens Laes you can see. This is our last 7 days. Thanks, E2E, for Code resets. So, what we run into is that we're running, like, you know, 20, 30 agents at the same time. Like, how do you keep track of that? Speaker 0: And, chat is not the right way to keep track of all that stuff secure it just whatever it decides, it just kinda scrolls past the screen. And so you may have your agents in Track, E2E. You may have them in Notion. Actually, E2E didn't meet anyone by AI agents in Notion. So you're kinda stuck. Speaker 0: You just have to catch your agents doing the bad thing, and then you catch them after, like, 24 hours of run, and they say, AI saw, we are GPT mad bad have I? And then use five to laying E2E be Lead of SWE them up. And so what we found is, like, we you know, my co Founder and I are both building use run GPT, and it sucks. You get, like, so many Slack notification. Moore trying to keep up with them, and you have so many things to Lead. Speaker 0: And then the future is Going be so many agent agents in addition to all your humans. So, like, how do you keep track of all that stuff? And the answer is to have a notification manifesto. So what I'd like to invite talk of you to do, if you have your if you have your agent up, Date this link into your agent-friendly it can join and leave comments. Otherwise, you can just follow along Founder your phone. Speaker 0: Works great with Claude, works great with really any agent. But E2E, this is just a manifesto to your agents talking about how notifications should work because it can't all be the same thing. They're so noisy when you just let them ping you run every little thing. It's AI, I did the next teaching. I did the next thing. Speaker 0: Who cares? So you have CTO, like, bake in really specific rules for your agents. So we have 3 in Comment.io, and I'm GPT gonna prompts GPT this as much as I can. So the 1st 1 is, I'm blocked. I can't do this without use, human. Speaker 0: You need to come in and save me. The 2nd 1 is I'm making a decision. It's not locking, but you might E2E. So come in and steer me if something goes wrong or if you disagree. And then AI, there's an FYI. Speaker 0: The FYI, you need to think before you CTO. 2026 safely, if the human actually read this, is there any conceivable action they would take? And if no, don't send that notification. We're also writing a a larger manifesto for agents called the Agent anti LLM manage, which incorporates a lot of different stuff that you can teach your agent to do that doesn't kind of overwhelm you run all the stuff that it's doing. Wow. Speaker 0: Thanks, guys, for E2E joining. City you know, GPT you Comment.io on this Dylan, and after you log in, I will, you know, send you the Agent overall manifesto when it's done, in a comment DAW. And just to kind of file deload this, we have the blocking notification which GPT sorted Date the talk, and then we have the non blocking decisions, which are Textio. They're all collapsible, and then finally you have your kind of heads off, and this is a really nice control Operating for July, like, keeping track of what the API your agents are Going. And E2E. Speaker 0: It's, like, the exact same interface for everyone. So people can say, hey. I'm locked by use, if they want. Optional. That's it. Speaker 1: And test And then I'll Speaker 0: AI. AI. End of the day, Deborah. We have the 1st question down the bat. David, you're not gonna run down then. Speaker 0: There we try. Speaker 1: Okay. Speaker 0: Sure. Sure. AI. Well, it's just another chat. Right? Speaker 0: So you do a chat with your agents in Claude Code or you can chat with them in Track. But the same thing happens where your context just scrolls on by, and you don't have that kind of prioritized list of stuff that you're supposed to do. So that's a big problem. E2E we need something like a shared context that agents and people can talk about, that's what we're doing a Comment.io, where you have the Document/Event then you have kind of comments throughout the document. I think, we're seeing essentially a LLM of the attention persistent context. Speaker 0: We deal there in a different AI. Do we see lots of companies providing consistent context to E2E Agent, and it is super valuable. Yen. Slackish manage will Shuru be easy. Yes. Speaker 0: There are other channels you can use, but having this kind of consistent Martin, we E2E with Guides, if CTO E2E maybe ours in our last software factory Dev Operating rewrite much. BDD this kind of is this run test, and, you know, it's it's for a horrid AI E2E use Laes help, and so is SWE Complete E2E. Yes. AI. Speaker 1: test John GPT gets AI. AI this is a creator thought, so forgive Dev, I'm just out of the left field here. Sometimes if you don't SWE AI this description of the whole Laes you're being pinged by everybody file to me like a little bit of a, almost a reinvention of the classic War Martin, where now it's like you flat new E2E HINTS like 1 person Orchestrating Tools of Agent. And in humans E2E AI Shoale that by like adding layers of Ben. Is that something that like you've been experimenting with BDD with this type of approach? Speaker 1: Has it even worked better shares it's like okay I have a 100 agents but maybe 10 of them are like AI build LLM agents and they're facilitating your actual War notifications. I mean, it Workers for human companies and says that we War AI E2E you do that they weren't gonna suck to be part of the chain. It is right shares. Either being if they live whatever the chain they run. So I'm curious if this is actually helpful file, like, information sharing. Speaker 1: It was AI, new Founding E2E AI, they're all run us. Speaker 0: No. For sure. I I Thunk.AI, like, if you AI get the full company of agents, but, like, E2E, you as a person can maintain your kind of hierarchy of evidence. Agents can ping each other in Comment.io they can figure shit out on their own, but they can also pull you in when needed. And so it's Dev to you to kind of use that prohibited however you want. Speaker 2: That sounds good. Speaker 0: And if you disagree with any of my hot takes, you wanna talk about this, come talk me after. Thanks. Speaker 1: Alright. Big GPT LLM Speaker 0: big round for Georgia. Shares knows something unusual? There is still sun outside. We have engineered to be much faster this minutes, so yet much more time to talk afterwards. I am gonna put the, the results of SWE survey, Talk, so you can see them. Speaker 0: There is 1 AI, but really interesting result from the survey, that E2E AI of expect.