Transcription: OpenAgents Episode 123 - GitHub Issues to Pull Requests

Source: https://x.com/OpenAgentsInc/status/1828629873063014755 Wiki source: https://raw.githubusercontent.com/wiki/OpenAgentsInc/openagents/Video-Series.md Media title: OpenAgents - Episode 123: GitHub Issues to Pull Requests We solve multiple GitHub... Upload date: 20240828 Transcription model: gpt-4o-mini-transcribe Generated at: 2026-06-01T16:51:08Z

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

[00:00] Speaker: Hey folks, I'm excited to show you my workflow these days. It's uh way better, faster, and cooler than working in code editors. We're gonna try to see how many of these issues we can resolve in this video, hopefully without touching an editor. So first let's do this, list open issues. So we have a coding agent, I call it a coding agent because it's an agent that has a bunch of tools hooked into the GitHub API. And there it listed my open issues. I'm gonna say pick one and solve it. Ordinarily I would pick one and kind of work with it and oversee it and stuff, but um let's just see what it does. All right, it shows issue 78 tool result shows malformed data as it seems to be straightforward. Let's start by creating a new branch. Creates a branch, issue 78. We can see the issues in the Git tool params. It's displaying complex value for non-string values. Let's just kind of take a tour of what it's doing here. So we have branch. Some of these are old. Let me delete the old ones. All right, so we got fixed tool result display. Do we have any changes there yet? Yeah, we do. 5 seconds ago. Let's take a look. And it's creating a pull request. Let's take a little spin. Did it make a pull? Yeah, it did. So fixed tool result display for complex values. This PR addresses issue 78. And it added a comment. As me, I might add a separate account for it later. A pull request has been created to address this issue. You can review the changes in PR 80. Update this thing, blah, blah, blah. And then Vercel will build it. And you can see the issue currently, which we're going to test in a second to see if it fixes. In post GitHub comment, it shows it. I'm not sure how many other tool results show this thing. But if it works, then we won't see that. And when our staging deploy finishes, we'll be able to see the same chat in our staging environment. So we'll compare it. Drum roll, please. While that's going, why don't I hold down Command, click over new chat to open up a second pane, and say, list open issues and pick one to solve, not 78. Let's just see if that can just run while we're working on this other one. Deployment has completed. Let's visit the preview. Yeah, we've got to log in. Or is this on our development database? Dang it. Well, let's do this. Let's go, say, openagentsinc v2 main. Make a demo comment on this issue so we can test if that worked. OK, we went from issue number raw data to issue number 78. OK, great. Solved. Let's go back to our original chat and say, you solved it. Can you merge and close? Actually, I don't think it can merge. But I'll do it. That's not a good name. Let's call it this. OK, that's the flow. You can stop the video here. That's the point. There was an issue in our code. I had made a GitHub issue for it. Pointed it at it. It was able to gather the needed context. I reviewed it. I merged it. That will be live on openagents.com in 90 seconds. You solved it. I merged it. Now please close the issue with a comment saying, we solved it. Now if I refresh the page, hopefully this... Let's check the status of our deployments. We're going to have to pay for an upgraded to Vercel thing here because we have all these queued. We're going to be generating a lot of these. What's the upper limit of how many of these agents I can have? What's the bottleneck now? I'm the bottleneck of reviewing. Maybe I'll train a reviewer agent. Let's check in on our other chat here. It thinks that it solved that HUD thing. Deployment failed. Oh, app layout is not the right layout file. I don't think it is. I don't think so. I don't want to open my editor. I don't want to look at it. Okay, so one issue down. Let's test this one. The build failed. Why did the build fail? You is not a function. I don't know what that means. Okay, so this is one where I'm going to want to test locally. So let's give that a try. And what is the name of the branch? Git checkout issue. And let's run this locally. It is. Use ref only works. Okay, so. 78 is good. Goodbye. Failed to build. One of our outstanding issues is having it post a comment on a failing build into the issue so it can kind of loop. We've got a job working so far, but it's not the right syntax or something. I apologize for the oversight. All right, let's do git pull again and see if that solves it. Reset HUD. Does it work? I'm going to see if I do this. It does, yeah. It works, but I want the button to be a ShadUI button. Look in components UI button TSX, I think. And just an icon. Like a rounded arrow thing for refresh, kind of. And put it bottom left. Talking to your code base, having it just do it, you review it, you ship it. I've got a problem with this tool. Result is always showing the loading icon. All right, so after this one, we're going to try to get fancy and do like multiple at a time. I've got to refresh the page so that fixes. Okay, let's see what that did. Oh, it doesn't refresh the chats pane. That's okay. Yeah, so that's important because if you have things go off the screen, sometimes you can't like get it. I don't know how to fix that other than refresh. Okay, that's exactly what I wanted. That beautiful little ShadUI button. V0, eat your heart out. Okay, so the pull request, this is good. Ship it. Looks great. I merged it. You can close the issue. We're at two issues solved, two PRs merged in, how long has this been going? 12 minutes. Should we keep going or should we like quit while we're ahead? Seven issues. Seven issues. Okay, let's try something crazy. This will probably go off the rails, but like you get the point, right? Did I open my editor at all? I don't think I opened my editor at all. Oh, yeah, for the local testing. Testing, blah, blah, blah. Did I write any code in the editor? No. In fact, I didn't even like use the editor at all except for the terminal. Like I could have just done all that with a terminal. Okay. You ready to go crazy? Okay, and by the way, we used a little bit less than $5 of credits for that. So average of $2.50 of compute per closed issue. Is that a good deal? I think so. We're going to make so much fucking money. Okay, let's do the next thing. Let's do something crazy. Like, I really need a larger monitor. I really need a larger monitor. When Chris raises a seed round, he's going to get the largest fucking monitor that exists in the entire world. Let's try four at a time. Let's try four at a time. I can't put that over there. I can't make my face go off screen. Oh, because I'm doing full screen. All right, let's just try this. List open issues and pick an issue to solve. Not number 59. What's it doing? I haven't refreshed the page. Okay, not 59 or 61. 53. Can your cursor do this? Can your cursor solve four fucking GitHub issues simultaneously? I can just go AFK. See you later. Holy shit. There is so much more. There's so much more. We haven't even put in the graph rag to this yet. We haven't even pulled back in our custom agent marketplace. Get some community collaboration here. You get reusable plugins. Pull in our lightning micropayment rev share. Guys, this is... It's like so huge. I don't know. I've only got two PRs. What is this? Crack the fucking whip. Yeah, that's probably an improvement, but I want this to use AI. Let's just do a few little things here. So for 82, PR 82, I want those to use AI. Stream text. Not regex stuff. Scrape this URL and implement. Not stream text. Not stream text. I want generate text. Oh, did you think our agents couldn't scrape docs URLs? Of course they can. Thank you, Firecrawl integration. Create a new branch. Issue 61 can fix non-user balance. I don't know how to test that. I don't know how to test that because I don't know what causes it. So that's maybe not the best one to demo. But I'll take a look at it. Pain user balance. Okay. Looks like you're only focusing on client side, but it saves none to database. So you need to look at more stuff. What's even the issue here? We closed the issue as it... Whoa, whoa, whoa, whoa, whoa. What's going on here? Add tool to list pull requests in a repo. It seems that the implementation has already been started, but there's an error. Okay, hang on, hang on, hang on. Which one is this? Issue 53. Some of these had comments already. And it closed... It closed this. Oh, I guess this was already pretty much done. Oh. Just something that I had forgotten to close earlier. That's nice. We're down to six. So I'll close this one. Done. Okay, so fix stop chat prevention.

[19:59] Speaker: Chat stop when it's pressed, but if stopped in the middle of a tool call, it could prevent new chats from going forward. Some of the other issues, blah, blah. Another odd behavior. OK, I guess we should test this. Let's go to main first. Yeah, now we've got an error, it looks like. Looks like the pain, some pain refactor is using window dot. I think that needs to be in. Well, I mean, should I fix that now? It's not blocking us. OK, what am I testing? Chat stop when pressed. Oh, I believe that is. OK, so the current behavior, let's do a test. The current behavior is if I stop in the middle of a tool call, open the readme and summarize in two sentences. If I stop it in the middle of a tool call, well, actually. It didn't do the tool call. Yeah, see, so it makes that duplicate thing. And then I think if I stop it in the middle of a tool call, it'll like break it further. Or will it? Yeah, it will. OK, so that's two wrong things. Let's see if this fixed one or the other or both. And this is issue 67. And let's try that again. We'll use the same thing here. Enter. I'm going to stop. No. It did not fix the duplicate thing. Like it still shows the previous message in the following one. I think you need to look at use chat core. OK. What else can we do? This says we've now addressed the NAN user credit balance more comprehensively, including both client side and server side fixtures. Oh, if I want to like get a second opinion. I can like pop over to Claude. I won't be using this very much longer. Describe how this handles NAN. Describe how this handles NAN. I could ask GPT-4 Mini. After calculating the cost in cents, there's an explicit check for NAN. OK. Cool. Cool. Seems good. See, I don't know how to test this because I don't know what was breaking. But this like seems like something I'm cool to merge. And why did this just the comment thing fails. No, that deployment failed. Why? Queued. Damn it. I could just YOLO. I could just YOLO. I could also consider it might be time to move off Vercel. If I need to wait two minutes per single build. Like what is an agent centric version of this look like? Maybe I just should just bring back the old days of like FTPing into servers. Agent FTP. What is this one then? Oh, let's see if this one worked. Pull request 82. So I asked it to implement an AI feature using just the docs. Like I'll be surprised if that one works. Let's see what it did. Generate text example. Yeah. Yeah. Oh, that looks good. Does it work? I don't know if it works. Uh, but let's find out why. The build is failing. Oh yeah. Window is not defined bullshit. I think that snuck in the main branch. OK. Here's what I can do. Let's go to the branch. The branch is here. Git checkout main. Git pull. Git checkout here. And if we do PNP build, that'll fail. Because It's not failing. It is failing. Here we go. Tell me where window is not defined at. That's annoying. It failed, but we're not going to tell you where it failed. But I think it's because I'm in my editor. I'm in my editor. Earlier, I think we got an error from here. Which I should be able to just move that into there. Let's see if that was it. That was it. Cool. So let's push that. That might mess up the rest of my builds. So, should I just like push all this stuff? OK, so 59. I'm going to not deploy these real quick. Not deploy this real quick. Let's go to 59. You can see that we've got like just a bit more to automate. It's going to be another minute. Should I just push it because I'm impatient and risk breaking prod? I'm sorry, folks. I'm sorry if I break your chat. I just I want to try it live. And I want that in the main branch so the rest can Now this thing that I can't test, but that looks good. I'm going to also merge this. Vercel, I'm doing a lot of waiting here. The bottleneck in my process right now is review. So check this out, people. So I'll do the rest of this review offline, but let's make this point. We shipped two pull requests for two separate issues that the codebase had in the first 12 or 13 minutes of this video. We were able to more or less productively have four different issues be worked on by four different sub-agents simultaneously. A little bit stressful with the context switch between them and juggling testing environments and things. One of those issues had already been done, so that got closed. So we went from nine issues down to six. Pretty sure we'll be going down further from that. I'll do more review offline. But the point is that the bottleneck here is no longer code. And the code that we have been shipping over the last 24 hours, it works. If you've been using open agents to this point and you try it now versus a week ago, you can see that it's way better. It's a bunch of bug fixes. And those bug fixes have been coded 95 plus percent by AI. The bottleneck is no longer code. The bottleneck is now infrastructure and review, both of which we may fully or more fully automate soon. See you soon.