Build Daily
Tinley Park · August 23, 2026Three OpenAI Engineers Shipped A Million Lines. Your Ten-Hour Agent Run Starts Here.
This video explains how to manage long-running AI agent sessions by using progressive context shaping rather than relying on static instructions. It details practical methods for maintaining current project state files and decision logs to keep agents focused on evolving goals.
The outline
Tap a timestamp to jump there in the video.
OpenAI's Million-Line Project
Three OpenAI engineers shipped a million-line internal product in a tenth of the usual time using long-running agent sessions.
The Context Problem
Long agent runs face context issues where giant instruction files crowd out tasks and become graveyards of stale rules.
OpenAI's Map Approach
OpenAI replaced giant manuals with a short map pointing to active execution plans, decision logs, and architecture maps.
Anthropic's Progress Files
Anthropic uses progress files as portable memory between sessions to record current state, completed work, and failed approaches.
Progressive Context Shaping
This practice involves changing the small set of instructions governing next steps as work produces new evidence along the way.
Opening Prompt Basics
A strong opening prompt defines results, materials, boundaries, and first moves but can only express what is known before work begins.
Directing Work vs Placing Orders
As projects produce new information, the job shifts from defending the original prompt to changing what the agent treats as current assignment.
Arise's External Plan Example
Arise solved agent confusion by moving plans outside the conversation window and rebuilding short plan messages from latest state on disk.
Current.md Workflow
Using a current.md file allows agents to read updated state at job start and update it after consequential decisions across different tools.
What Belongs in Project State
Only save changes that affect the target or definition of done; explanations of how you got there belong only in history.
Codex Benchmark Case Study
A long Codex project required changing state to stop an unbounded generation loop and focus on deepening evidence for the most valuable answers.
The million-line codebase built entirely by agents underscores that scale demands structured context, not just raw output. Without clean, reusable context files, even powerful agents lose direction and coherence. For solo builders, this means investing in systems that preserve clarity across long-running projects.