In the video, the hosts tackle a coding challenge by initially attempting to use Rust, but pivot to HTML and JavaScript after realizing Rust isn’t necessary, leveraging custom AI agents and specialized skills to automate and verify their workflow. They successfully build and test a web application, demonstrating effective collaboration and experimentation with AI-driven coding tools.
In the video, the hosts discuss their approach to a coding challenge using Rust. They begin by explaining how they captured the challenge details—taking a photo with their phones and transferring it to their computers. The initial step was to check if their machines were set up to run Rust and whether it was feasible to build a website using Rust. Upon discovering that Rust wasn’t installed, they followed the installation instructions and prepared a prompt for the coding contest, aiming to clarify the rules and enhance the instructions for better clarity.
The hosts then describe how they moved into planning mode, focusing on creating a detailed plan for the challenge. They considered the possibility of deploying the project but decided that getting it running locally was the first priority. During this process, the coding agent they used pointed out that Rust wasn’t necessary for the task, suggesting that HTML and JavaScript would suffice. This prompted one of the hosts to switch to a custom agent called “beast mode rebooted,” which enforced specific coding principles and used sub-agents for task management, despite recommendations against it.
The custom agent was configured to use a specific context and to operate in an experimental “autopilot” mode, similar to a loop in Copilot CLI. This allowed the agent to continuously test and iterate on the code, including opening localhost and taking screenshots to verify progress. Eventually, the agent completed the task, but there was some confusion about how to run the resulting project. After some troubleshooting, they managed to get it up and running.
A key part of their workflow involved using the front-end design skill from skills.sh, which one of the hosts had pre-installed. They mentioned that Theo recommended using Opus46 with the front-end design skill as an optimal approach for design tasks. Additionally, they utilized the agent browser skill to enable the agent to test the application in a browser, addressing a common issue where agents claim to have completed a task, but the result doesn’t actually work. This skill allowed the agent to verify that the application was functioning as intended.
To conclude, the hosts tested the application’s functionality by giving kudos to a user named John and observing the animation. They checked if the kudos persisted after reloading the page, which it did, indicating successful implementation. The hosts expressed satisfaction with the outcome and playfully asked if they had won the challenge, highlighting both the technical achievements and the collaborative, experimental nature of their live coding session.