{ "video": { "id": "UlFB6efYN5Q", "title": "A Song of Types and Agents - Roberto Stagi, Ratel", "duration": 856, "upload_date": null, "channel": "AI Engineer", "source": "AI Engineer", "event_format": "Online" }, "analysis": { "video_id": "UlFB6efYN5Q", "title": "A Song of Types and Agents - Roberto Stagi, Ratel", "one_liner": "Roberto Stagi argues that TypeScript overtook Python as GitHub's most-used language in August 2025 because coding agents made it the default output language, and that while models still train and serve on Python, the agentic application layer now belongs to TypeScript.", "summary": "The talk frames Python vs TypeScript as a war for the AI throne and claims TypeScript is winning the layer that matters now. The argument: AI moved up the stack from infrastructure to the application layer — 'applications that think' — and the application layer has been TypeScript's for years, so the surge of AI integrations lands on TypeScript rather than Python. Stagi points to GitHub's own reports crediting AI for Python's 2024 rise and TypeScript's August 2025 takeover, and attributes the flip to coding agents (Lovable, Claude Code, Cursor, Codex) becoming the default way to build, with TypeScript as their default output. He closes with concrete reasons to build agents in TypeScript — one language end to end, Zod as a single schema, the npm long tail, a fast-growing AI SDK ecosystem — while conceding training, research, GPU serving and inference stay Python's.", "key_points": [ "In 2024 Python became the most popular language on GitHub, and GitHub's report credited AI: 'AI leads Python to the top language.' In August 2025 TypeScript passed Python, and the report gave the same reason: 'AI leads TypeScript as the first language.'", "The variable that changed between 2024 and 2025 was coding agents growing up — Lovable, Claude Code, Cursor and Codex became the default way to build applications, and the default language they built in was TypeScript.", "Developer numbers surged in both years; in 2025 one new developer joined GitHub every second.", "AI stopped being something you train and became something you ship inside an application — and the application layer has been TypeScript's for a long time, not Python's.", "Anthropic acquired the Bun JavaScript runtime last December, cited as evidence an AI lab is investing in the JS/TS runtime layer.", "A self-reinforcing loop: more TypeScript applications feed the training of the next coding agents, so agent output quality in TypeScript should keep improving.", "TypeScript gives one language and one codebase for the agent loop, tools, backend service and UI, with Zod as a single schema defined once and reused in backend, model and UI; a Python stack typically splits into at least two services (FastAPI + Pydantic AI, plus a separate React app) with a contract and two type systems to keep in sync.", "The Vercel AI SDK went from 1.6 million to 15.1 million weekly downloads in one year — roughly 9–10x — as evidence the TypeScript AI ecosystem is surging.", "npm is framed as the richest package manager and deepest long tail — authentication, payments, UI, infra — which matters precisely because AI now has to integrate with the application layer.", "Prediction: the gap on the application layer will widen from here; models run on pip, agents ship on npm." ], "takeaways": [ "Keep training, research and GPU/inference work in Python — the speaker is explicit it isn't going away — but default to TypeScript for new agents and AI-powered applications.", "Collapse the two-service split: put the agent loop, tools, backend and UI in one TypeScript codebase instead of a Python service plus a separate React app with a hand-maintained contract.", "Use Zod to define a type once and reuse it across backend, model and UI rather than maintaining parallel type systems either side of a language boundary.", "Lean on npm's existing auth/payments/UI/infra packages for the integration work the application layer now demands.", "Factor coding-agent quality into language choice: agents are best at TypeScript today and that gap is expected to widen, so overlooking TypeScript risks falling behind." ], "topics": [ "typescript", "python", "agents", "coding-agents", "application-layer", "developer-ecosystem", "type-safety", "npm" ], "tools": [ "TypeScript", "Python", "JavaScript", "GitHub", "ChatGPT", "Claude Code", "Cursor", "Codex", "Lovable (transcribed as 'Lava')", "Anthropic", "Bun", "npm", "pip", "Zod", "FastAPI", "Pydantic AI", "React", "Vue", "Vercel AI SDK (transcribed as 'Versatile AI SDK')", "Ratel (a context layer for AI agents)", "AI Tinkerers (transcribed as \"AI's Pratic\")" ], "quotes": [ { "text": "AI stopped being something that you train, and it started being something that you ship inside your application.", "at": "02:57", "url": "https://www.youtube.com/watch?v=UlFB6efYN5Q&t=177s" }, { "text": "since every new app pretty much every new app is an agent today because they ship these AI and agentic capabilities, they are hungry to embed AI inside themselves. The demand to have more AI integrations, more and more AI integrations doesn't fall on through Python. It falls on TypeScript.", "at": "06:19", "url": "https://www.youtube.com/watch?v=UlFB6efYN5Q&t=379s" }, { "text": "Jeff Atwood said, \"Any application that can be written in JavaScript will eventually be written in JavaScript.\"", "at": "12:12", "url": "https://www.youtube.com/watch?v=UlFB6efYN5Q&t=732s" }, { "text": "the model can still run on pip. But the agents, which is the application layer today, so the agent that called the models will probably ship on NPM.", "at": "13:05", "url": "https://www.youtube.com/watch?v=UlFB6efYN5Q&t=785s" } ], "words": 2081 }, "summary_url": "/#UlFB6efYN5Q", "transcript": { "html": "/transcripts/UlFB6efYN5Q.html", "txt": "/transcripts/UlFB6efYN5Q.txt", "vtt": "/transcripts/UlFB6efYN5Q.vtt" } }