Transcription: OpenAgents Episode 138 - Year One Recap

Source: https://x.com/OpenAgentsInc/status/1855464290771247126 Wiki source: https://raw.githubusercontent.com/wiki/OpenAgentsInc/openagents/Video-Series.md Media title: OpenAgents - Episode 138: Year One Recap We recap the first year of OpenAgents an... Upload date: 20241110 Transcription model: gpt-4o-mini-transcribe Generated at: 2026-06-01T15:23:05Z

Machine-generated transcript. Review speaker labels and wording before using this as quote-grade source material.

[00:00] Speaker: Hard to believe we're celebrating one year of this series. Year one, first of many. We're at 137 episodes, and you know, you might think that's a lot, but we're trying to catch up to Handmade Hero here, this project from Casey. You know, they got they got they got they got 667 episodes. So we're just we're just getting started here at Open Agents. But I wanted to review what we did for the last year for those just tuning in or curious what we've been up to. And let's start by asking Open Agents what we've been up to. So scrape the video series link and summarize the topics here in two paragraphs. This is our pro interface. We're gonna start opening this up to the public quite soon. But yeah, our video series so far has documented the development of an open source AI agents platform, spanning 137 episodes from initial setup to advanced features. The early episodes cover foundational work, including tech stack selection, Laravel implementation, and basic agent principles. The series then progresses through key developments like PDF chat capabilities, RAG implementation, and the creation of, later called AutoDev, a GitHub integrated coding agent. Later episodes showcase more advanced features, including plugin systems, Lightning network integration for payments, agent node graphs, and various UI improvements. Recent episodes focus on AutoDev capabilities, code base indexing, and the development of OpenPress. The series demonstrates an evolution from basic agent functionality to a sophisticated platform incorporating AI coding assistance, payment systems, and developer tools with a strong emphasis on open source. Okay, so that's all more or less accurate. I would say one kind of problem or with the problem with that rosy explanation there is that all of that is kind of split across a few different code bases. Our agent store that we launched four or five months ago ish is in the V1 code base. That is not no longer actively maintained. We'll go into some of the lessons learned from that shortly. The version two interface, sort of our developer focused auto dev interface for interacting with GitHub code bases, that's still live at stage2.openAgents.com. I've been using this every day for the last number of months until this past week, starting to move that over to our version three pro interface, which we're going to start rolling out more publicly. This has supports for teams, more tools, more repos, latest models, and we're going to start pulling back in some of the agent store and incentivized plugin stuff from our version one interface. So let's step through a few of the lessons that we learned. So I'm going to say lesson number one is kind of revenue matters most. It's kind of related to two and three here, which is before we get too creative on protocol things like our version one interface, we did a whole bunch of creative stuff around Noster and we kind of think got a little bit ahead of ourselves because what we ended up running into was actually a similar problem to what we ran into with GPU Utopia before this, which is we had a bunch of supply lined up. Like when we did our GPU Utopia, our sort of decentralized GPU compute network using lightning, we had a whole bunch of sellers. We had a whole bunch of compute sellers happily come by and want to sell compute, contributing it to our open AI compatible API endpoint that supported inference, fine tuning, embedding, and image generation. The problem there was that we just didn't have any real buy side demand lined up. It was difficult at that time to be competing against sort of the sort of like VC subsidized race to the bottom in terms of infrastructure prices back when we were like specifically trying to pitch this API. And we ran into like pretty much the same problem with the version one of our agent store, which is it was great that we had the ability technically to have people be able to contribute agent plugins to this marketplace and have, you know, some usage on the front end and have the lightning payouts go to the people that were building the plugin. The problem was that we were kind of pitching that at this sort of like consumer audience where it wasn't really clear what they would want to use it for. We had kind of hoped that sort of like market forces would figure it out, and market forces did not really figure it out. And maybe that was just mismanagement on our part or not doing or saying the right things. But the kind of like lesson that we learned from there is like, okay, we really need to start by focusing on what is the first use case for which people are actually willing to put down a credit card or, you know, chunk of Bitcoin and pay for real usage of something. Our first kind of calculation that that would be coding agents. So we built our version two coding agents just to be like a really good workflow for interacting with a GitHub code base and being able to have, you know, an agent loop through and do meaningful things over like a 90 second time window where people are able to be more productive as a developer here. And that was okay. We had a few, you know, developers using this happily. I have been using it happily and heavily. But the real success in generating revenue came only recently when we started shifting to business customers, people that have been either either having challenges hiring or they've got basically pain points that are not in the tens of dollars per month or the hundreds of dollars per month, but they're in the thousands of dollars per month. So we've been using our sort of agents, agent interface, the auto dev interface, and this new pro interface sort of in this applied format directly to a few customers. Call it a private beta, whatever. And that has addressed the thing that we had not had until this point, which is what is the use case that people are willing to pay money for? Now that we have that, now that we have an idea of like, what is the profile of the person, the company that wants to pay thousands of dollars a month for some sort of compound AI system that solves problems for them? Now we can reason about, okay, how do we extend that into things like, you know, some of our customers want more than just the current integrations that we offer. They might want something that goes to their Slack or just other types of things that you can imagine that this would be absolutely perfect for the whole agent plugin idea that we like previously spec down built now to combine that in. So anyway, we're, we're, we'll get to that in terms of like what we're doing next. One of the other lessons was that the simplicity of the tech stack does not equal simplicity of development. I have said even recently that I've gotten tired of certain of the like, you know, the big JavaScript and I've, I've been really kind of taken by the, the sort of swan song of, of, you know, siren song of HTMX and doing things simply and getting away from like big JavaScript and stuff. And, you know, you know, started V1 with Laravel, did V2 with Vercel and Next.js and React. V3, you know, tried a bunch of different ways to use Go or HTMX or Laravel without a bunch of like crazy JavaScript. And I just came to the realization that it's really, really, really difficult to be more productive than what the, you know, VC backed companies are spending a lot of time, energy, and resources to put into your hands in terms of specifically AI tooling. The Vercel, for example, the Vercel AI SDK that powers the sort of chat and tool use of our version 2 and now version 3 interfaces. It's, it's hard to beat. I've tried to recreate that in multiple different ways. I've got it like working decently on the front end, but then the back end, like there's just, they solve a whole lot of stuff. It's really hard to get away from. So it's like, all right, one lesson is just like, stop being cute. We can get in the future when we want to do things in a different way and use, I don't know, other stacks. But like, it's really stop being cute about tech. Just do what's necessary to get the product out. And that kind of gets to like, yeah, product before protocol or like, you know, letting my, my inner computer nerd that wants to do like the cutting edge that no one else is doing or like be HTMX to be funnily contrarian. Like I, I want to do that. I spiritually want to be doing that, but like, realistically, we got to ship product that works. Uh, if you know, there's a new AI integration that comes out and Vercel adds it the next day instead of needing to build that myself. It's just like the practicality of that has won out. So just like a principle that we need to hold. Um, so some takeaways here are our core thesis is intact. I expressed in the opening video of this series, like, um, we now agree basically with Sam Altman. This was done like the day after the open AI dev day presentation a year plus ago where, um, they're all talking about that. The future was going to be agents. I said, you know, I absolutely agree with that, that, um, this agents are the things that businesses are going to be paying for. People are going to be paying for. And to take that one step further, I believe that every person, every business, every family, every street corner, every device is going to have one or more agents, AI agents, basically compound AI systems working either on your behalf or as part of the system or as part of the device or part of the physical location. Agents will be everywhere. Do you want these massive closed source corporations with poor incentives, with every incentive to get their tendrils around the government, scaremonger about AI safety and try to erect barriers to keep out startups? Like we cannot let that happen this time. It's the stakes are too high to mess with that