Transcription
You are using Cloud Code incorrectly. While you are typing in one prompt after another, I am currently running 16 agents simultaneously, and after that, the next 82 are already scheduled. They orchestrate and control each other before anything even reaches me. My name is Loris, and I work full-time with AI to program AI. And I can truly tell you, this new Cloud Code feature is an absolute game-changer. Before we begin, I would really recommend subscribing to my channel if you don't want to miss anything from the AI world. But let's start from the beginning. Recently, Cloud Opus 4.8 was released, and yes, it's just a new model. The Trust Me Bro benchmarks are, as always, a bit larger than before, and we all love it when numbers get bigger. Two other things are interesting about this release. There's now a new setting called Ultra Code and the feature Dynamic Workflows. Ultra Code combines X High Reasoning and automatic workflow orchestration. A workflow is a small script that Cloud writes for itself for its task and that deterministically controls a whole swarm of sub-agents in the background. This scales massively, and Cloud's context remains much freer. You can run up to 1000 agents per run, with 16 in parallel. And you know what? I once made a video about the JavaScript Runtime Ban. Ban is, of course, a relatively large project, which until now was implemented in the programming language Sig. But Jared Samner, the lead developer of Ban, who now also works for Anthropic, simply rewrote the entire codebase in Rust with the help of a Dynamic Workflow. The PR adds over a million lines, and 750,000 of those are in Rust. It was also merged by the way, so B is now in Rust, thanks to this new feature. You can manually trigger workflows by telling Cloud to create one. Or you switch to the Ultra Code setting in Cloud Code with the command /effort. Here, Cloud automatically decides when to create a workflow and when not. And I have to say, it works really well. And because it's a script, you can of course also save and reuse a workflow. But what does such a script actually look like? So, it looks like this. This script is a code reviewer that looks for bugs. You can think of it as a kind of conductor. It doesn't do the work itself, but distributes it to many sub-agents. In this script, it works so that first Finder agents look for bugs, one for each topic, and then Skeptic agents try to refute each bug. Only the bugs that survive the skeptics are returned as results. The individual components within the script are then first a meta-brief, which describes what this script actually does and divides it into phases. Then there is a schema for the bug finders in Finding, which describes in what form the agents must respond. The same thing exists for the skeptics, here called Verdict. Each skeptic must clearly state at the end whether they were able to refute the bug or not. In Dimensions, perspectives are defined under which a repo should be viewed. For example, Security, Concurrency, or Correctness. The actual process is in the Pipeline. There, each entry from the Dimensions is sent through two stages. In Stage 1, the Finder agents are started, and in Stage 2, the results here, referred to as R, are checked by the Skeptic agents with the prompt "You are a skeptic of D. Try to refute this finding." In this script, it's always Best of Three. So, three skeptics try to refute a bug. If two or more skeptics agree with the bug, it survives the pipeline, and in the end, you only have confirmed bugs, which are then returned by the script. So, scripts are actually totally simple, but oops, no, but we'll test this on a real project. [music] I thought it would be cool to find a real issue in a widely used Git repo. I did some research and found a relatively new, not yet incredibly large, but still cool repo. And there I found LightPs. LightPs is a local tool that can read PDFs and extract their text in a structured way. For example, if you want to pass the content of a PDF to your AI agent, you can read out this entire content with LightPs and pass it directly to an AI. The initial commit of LightPs was 4 months ago, so the repository is still relatively young. Nevertheless, it already has 9000 Git stars. So, it is actually used by some people. It is also actively maintained, as the last commit was not too long ago, for example. And now let's become an open-source contributor. To do this, I open Cloud Code and with /effort I now switch to the Ultra Code setting. In the subsequent prompt, I then tell Cloud to clone the repo and find real [clears throat] bugs, but also to verify each bug so that only reproducible bugs remain. This demand for accuracy or reproduction implicitly tells Cloud to start a workflow, which it then does. While the workflow is running, we can even look at the whole thing with /workflows. Um, while the workflow is running, we can look at the whole thing with /workflows. Okay, in Cloud Code, you always have to wait until the current prompt has finished. Only then is the next command accepted, but then we get into this cool view where we see the individual phases on the left, can go to the agents within a phase, and can check individually what each agent is doing. It's actually extremely entertaining to watch. Here we are currently starting 14 Finder agents that are actively searching for bugs, and then another 99 Skeptics who then check these bugs, and finally another agent who then ranks the bugs. So, it works roughly like our example from before. When everything has finished, we have to manually leave the workflow view. If you don't do that, Cloud simply won't proceed here anymore. This would be, for example, an issue that I would like to report, but no, Cloud Code is not open source. Well, we can get out of this situation with Escape, then we are back in the Cloud session. Cloud then summarizes the results of the workflow for us here and presents all the bugs to us. There are six top findings, but you shouldn't just blindly accept them. The problem is, there's an incredible amount of GitHub issue slop. That simply means that people let AI bots search GitHub, scan repos for bugs, and then open issues. Mostly, it happens completely automatically. Of course, that sounds good at first – people get their open-source contributions fully automated, and repo developers are shown their bugs so they can be improved. But the problem is, the vast majority of them are not real bugs, it's issue slop, and so only the issues of open-source repos get cluttered, which leads to developers finding even fewer bugs in the end because they get lost in all the slop. But a token maxer is not a slop maxer. We will now manually review all the bugs again. Okay, I was able to confirm four out of six bugs with my own eyes, but nothing security-critical among them. One of the bugs, for example, was that if you accidentally specify an extremely large page number, LightPs tries to write every single page number into a list and consumes up to 17 GB of RAM. And you know, RAM is expensive. Okay, the bugs are not world-shattering, but bugs are bugs. Now we can simply ask Cloud to write the issue messages for each bug directly for us. But wait, what a cool idea Cloud has come up with again! Cloud has direct access via the GitHub CLI and can also create the issues in the repo itself. And that's what it did, and now we are officially open-source contributors and have done something good for a change. And do you remember the agents from the beginning? This video was researched using roughly the same architecture as our bug hunter. For this deep research run, agents performed a parallel web search, researched facts, and provided the sources directly. Of course, I then checked all the facts myself, but I actually didn't receive any false facts. Everything was correct. This workflow even found an error of mine that I had written into the prompt. I thought Ultra Code was an Afel, but no, it's a setting. Sorry. Well, what do you think of Ultra Code and Workflows? Will you integrate it into your work, or do you simply consider it pure token maxing? Write it in the comments. And that's it for today, my children. If you are interested in AI and AI programming, then you should watch this video now. Thanks for watching and see you next time.