The Agent So Good It One-Shotted My App

Detailed Summary

Introduction to LLM Challenges and Deep Agent Desktop (0:00 - 1:02)

• The presenter highlights the significant hardware requirements and potential quality loss when running large LLMs locally for coding.
• Introduces Deep Agent Desktop, an AI-first, agent-first fork of VS Code, as a solution for accessing powerful LLMs without local hardware constraints.
• Deep Agent Desktop offers access to various models, including popular frontier models like Claude Sonnet 4.5 and open-source options like Kimmy K20950.
• Explains that Deep Agent is part of Abacus AI's suite of tools, which includes image and video generation, all accessible through a single monthly fee.

Deep Agent Desktop Features and Application Goal (1:02 - 2:48)

• Deep Agent Desktop includes AI-first features like chat mode and a command-line interface (CLI) for agent interaction.
• The presenter aims to build a practical, real-world application: a link management tool similar to Bit.ly, rather than a simple portfolio website.
• A key trick for effective agent use is demonstrated: asking an LLM (GPT5 in this case) to generate a detailed, one-shot prompt for the coding agent.
• The desired application will use Next.js with a database, specifically excluding cache and Vercel's ClickHouse analytics tool for simplicity.

Prompt Generation and Sonnet 4.5 Test (2:48 - 4:57)

• The LLM generates a comprehensive prompt detailing the tech stack, core functionality (link management, basic analytics, authentication), data model, API layer, validation rules, and environmental variables.
• The presenter initiates the process with Claude Sonnet 4.5 in Deep Agent Desktop, providing the generated prompt.
• Sonnet 4.5 successfully processes the request, and after a brief pause for Docker to be enabled (for the database), it produces a fully functional application.
• The application includes working authentication and a link creation feature, demonstrating a successful redirect, which greatly impresses the presenter.

Kimmy K2 Test and Troubleshooting (4:57 - 9:06)

• The presenter sets up a new, blank folder for the Kimmy K2 test within Deep Agent Desktop, selecting Kimmy as the agent.
• Kimmy K2 begins generating files rapidly, leveraging remote servers for processing power, which is a benefit of the platform.
• However, Kimmy K2 fails to run the application automatically, requiring manual intervention for npm install.
• The agent creates an incorrect directory structure, leading to invalid package name errors.
• Further troubleshooting reveals a hallucination in the code: types.reactd instead of type /reactdom.
• Despite fixing the package name, the application still encounters an unhandled runtime error (Input is not defined), indicating further issues.
• The presenter notes that this back-and-forth troubleshooting is a common occurrence with AI coding tools, unlike the seamless experience with Sonnet 4.5.

Comparison and Deep Agent Platform Overview (9:06 - 10:48)

• Based on the experiment, Claude Sonnet 4.5 is deemed superior to Kimmy K2 for this specific task, highlighting the value of having access to multiple models.
• The presenter emphasizes that Deep Agent is more than just the Desktop application; it also offers browser-based agents with extensive capabilities.
• Deep Agent can use your computer, integrate with services like Gmail, Google Workspaces, Slack, and Jira for automation, and even post tweets.
• Additional features include deep research, presentation creation, charts, docs, reports, video generation (including lip-sync), and image generation.
• The presenter concludes by noting that the Chat LLM Teams subscription combines all these tools, offering significant cost savings compared to individual subscriptions.

Notable Quotes

"You know, I'm no stranger to running LLMs locally, especially for coding purposes, but you need a lot of hardware like I have here, like I have over here."
— Alex Ziskind

"This is called Deep Agent Desktop. And those of you with a keen eye might recognize this. It's a fork of VS Code, but this is AI first and agent first."
— Alex Ziskind

"Fully functional application with authentication that works. This is kind of why people like using Cloud Sonnet 4.5. Amazing."
— Alex Ziskind

"I cannot believe where we are right now. This is crazy. I mean, every time I do this, it's like some parts are pretty amazing when it works. Like when it works immediately."
— Alex Ziskind

"Can we judge the difference between Claude Sonnet 4.5 and Kimmy at this point? Well, based on this experiment, yes, we can."
— Alex Ziskind

"Deep agent actually has the ability to use your computer. It also integrates with Gmail, Google Workspaces, Slack, Jira, so it can automate all those processes for you."
— Alex Ziskind