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

What this video covers

Jones examines how a small engineering team can use coding agents across very long sessions without expecting a single prompt to contain everything the system needs. The key practice is progressive context shaping: humans repeatedly improve the information an agent receives as the work evolves, instead of treating context as a static input prepared once at the beginning.

He separates that context into durable instructions, an accurate account of the current state, a navigable map of the codebase and a history of decisions or failed approaches. These layers help the agent recover after context compression, distinguish stable rules from temporary details and find the parts of a large project that matter for the next step.

Checkpoints remain a human responsibility. Jones recommends asking agents to plan, inspect their own work, run verification and leave the repository in a state another agent can understand. The advantage of longer runs comes from better steering and preserved project knowledge, not from removing people from architectural judgment or quality control.