Three OpenAI Engineers Shipped A Million Lines. Your Ten-Hour Agent Run Starts Here.

AI agent context files are what keep a long agent run on track once your opening prompt goes stale. Here is how OpenAI, Anthropic, and Arize actually structure them, and the four files you can copy. Grab the Working Context Starter Kit on my Substack: https://natesnewsletter.substack.com/... My Links 🔗 👉🏻 Newsletter: https://natesnewsletter.substack.com/ 👉🏻 X: https://x.com/natebjones 👉🏻 TikTok:   / nate.b.jones   👉🏻 Instagram:   / nate.b.jones   What's really happening inside a six hour agent run? The common story is that a bigger context window solves this, but the real question is which decision the agent treats as current. In this video, I share the inside scoop on progressive context shaping: Why one giant instruction file becomes a graveyard of stale rules How to change an agent's direction without restarting the project What four kinds of context you should keep separate Where OpenAI, Anthropic, and Arize put the current plan More capable agents extend the reach of your judgment, and they extend the reach of an outdated judgment exactly as far. Chapters: 00:00 The million line project and what the headline hides 02:43 OpenAI swaps the giant manual for a short map 03:21 Anthropic's progress file as portable memory 04:41 Progressive context shaping, defined 05:36 Why the opening prompt cannot stay in charge 07:04 Arize, 27 model calls, and the plan moved to disk 10:21 My Codex benchmark and the instruction that went stale 13:07 The four kinds of context worth separating 16:14 400,000 Claude Code sessions and the 70/80 split 17:15 Symphony, three to five sessions, and the ticket board 18:35 The four things to give an agent at the start 21:02 Why context window size is not the capability Listen to this video as a podcast. Spotify: https://open.spotify.com/show/0gkFdjd... Apple Podcasts: https://podcasts.apple.com/us/podcast...