# How I code with AI agents, without being 'technical' Source: https://www.youtube.com/watch?v=snF5eGKoiJI Recap page: https://rapidrecap.app/video/snF5eGKoiJI Generated: 2026-01-07T22:32:28.638+00:00 --- ## Quick Overview Coding with AI agents without being technical is achievable by learning the underlying systems (like Webflow, Zapier, and Airtable's roles) rather than just memorizing syntax, as demonstrated by Ben Tossell who spent 3 billion tokens in four months using agents to write code he couldn't write himself. **Key Points:** - Ben Tossell spent 3 billion tokens in four months watching an agent write code he couldn't write himself, proving non-technical coding with agents is possible. - Tossell rejects the 'vibe-coder' or 'no-code' labels, suggesting the process involves learning the underlying systems of how code works, where things fail, and where they succeed. - He lists projects shipped using agents, including a personal site revamped as a terminal CLI tool, a social tracker for 'Factory' mentions, and custom CLIs like a Pylon CLI. - His working style involves using a CLI exclusively, feeding context to the model, switching to 'spec mode' to develop a plan, linking docs/GitHub repos, and then letting an Opus 4.5 agent run with high autonomy. - He emphasizes learning from others by studying how engineers like Peter Steinberger and Mario build systems, noting that his own learning process is about identifying and fixing gaps in his understanding of the system. - The author highlights the benefit of tools like Droid (which he uses because he works at Factory) for providing agent-specific setup instructions in an `agents.md` file, which simplifies setup across repos. - The key takeaway for learning is to embrace the 'fail forward' mentality, realizing that mistakes (hiccups, bugs, or issues) are opportunities to learn, and that the feedback loop is much quicker than traditional learning. ![Screenshot at 00:09: The video opens by displaying Ben Tossell's viral X thread titled, 'How I code with agents, without being 'technical'', which serves as the primary source material for the discussion.](https://ss.rapidrecap.app/screens/snF5eGKoiJI/00-00-09.jpg) **Context:** The video discusses the feasibility and methodology of coding complex software using AI agents, specifically referencing the experiences and insights of Ben Tossell, who authored a highly viewed thread on X (formerly Twitter) titled, 'How I code with agents, without being 'technical''. Tossell details his intensive use of AI agents (consuming 3 billion tokens in four months) to generate code and build projects, contrasting this new approach with traditional coding education and the limitations of previous no-code tools. ## Detailed Analysis The speaker analyzes Ben Tossell's experience of coding using AI agents without traditional technical skills, primarily by consuming 3 billion tokens over four months watching an agent write code. Tossell rejects terms like 'vibe-coder' and emphasizes that this new technical class requires understanding the underlying systems (like Webflow, Zapier, and Airtable) rather than just memorizing syntax. He details projects shipped, including a CLI-like personal site and custom CLIs, which he built by interacting with an agent (specifically Droid, which is model-agnostic). His workflow involves giving the agent context, switching to 'spec mode' for planning, linking GitHub repos, and then running Opus 4.5 with high autonomy while actively monitoring and guiding it. A key lesson is the 'Fail Forward' mentality: since the agent handles the basic debugging, one should feel comfortable asking 'silly' questions and iterating quickly, which is faster than traditional learning. He notes that learning from others' open-source contributions and adopting structured documentation like `agents.md` files accelerates this process, ultimately shifting the focus from learning syntax to understanding system architecture and leveraging AI as an extremely capable general agent. ### Introduction to AI Agent Coding - Can you code without being technical? - Tossell spent 3 billion tokens in four months using an agent to write code he couldn't write himself - Tossell rejects 'vibe-coder' and 'no-code' labels. ### What I've Actually Shipped - Personal Site (revamped as a terminal CLI tool) - Feed (simple social tracker for company mentions) - Factory Wrapped (first version of wrapped product) - Custom CLIs (like a Pylon CLI) - Crypto tracker - Droidmas (12 experiments) - AI-directed video demo system - A Telegram bot powered by Droid Exec. ### How I Actually Work - Uses CLI exclusively over web interfaces - If an idea arises, spins up a new project in Droid (Factory's CLI) - Generally talks to the model for context, then switches to spec mode for planning. ### Spec Mode & Agent Interaction - In spec mode, he questions things; links docs/GitHub repos to the agent - Lets Opus 4.5 run with high autonomy, watching the stream for errors and guiding the agent if needed. ### Learning Process & Mindset - Learning is about identifying gaps in understanding systems built with code, not just syntax - Mentions learning from Peter Steinberger's simplicity and Mario's MCP posts on CLIs vs. MCPs. ### Why This Is Different - The shift from writing code from scratch to learning how to work with an AI agent (prompting, context, etc.) - Old no-code abstraction was drag-and-drop (Webflow, Zapier, Airtable); new abstraction is more programmable. ### Fail Forward Philosophy - Learning code means building ahead of capability and failing forward - Encourages asking 'silly' questions because the AI handles basic debugging, which accelerates learning. ![Screenshot at 00:09: The video opens by displaying Ben Tossell's viral X thread titled, 'How I code with agents, without being 'technical'', which serves as the primary source material for the discussion.](https://ss.rapidrecap.app/screens/snF5eGKoiJI/00-00-09.jpg) ![Screenshot at 00:57: Introductory segment showing the podcast branding \('The Startup Ideas Podcast'\) overlaid on a bright orange background with milk carton graphics.](https://ss.rapidrecap.app/screens/snF5eGKoiJI/00-00-57.jpg) ![Screenshot at 00:10: A close-up view of the X thread showing Tossell's usage statistics \(1.6K sessions, 3 Billion tokens\) and the title of his post.](https://ss.rapidrecap.app/screens/snF5eGKoiJI/00-00-10.jpg) ![Screenshot at 04:09: The screen transitions to a blank interface with a search bar, demonstrating the user interacting with an AI environment by typing a query \('Factory's CLI'\).](https://ss.rapidrecap.app/screens/snF5eGKoiJI/00-04-09.jpg) ![Screenshot at 04:16: A demonstration of the 'Factory' website, showing a section titled 'Command Line: Automation Across the SDLC' with example CLI code snippets.](https://ss.rapidrecap.app/screens/snF5eGKoiJI/00-04-16.jpg)