Your AI can write a thousand questions when you needed the 50 best answers. How to keep a long run pointed at what you want now, even as the goal shifts

The principles, technical limits, and shared team methods behind long-running agents, with a companion of worksheets and prompts.

TL;DR

  • AI systems are increasingly capable of long-running tasks, investigating problems for hours or days and self-correcting.
  • Major AI labs like OpenAI, Anthropic, and Google are reporting experiments with extended AI sessions for tasks such as code management, operating system development, and scientific research.
  • This capability allows humans to delegate more complex work and focus on strategic decisions, but requires AI to recognize progress, retain learned information, and adapt to changing goals.
  • Future developments need to address how AI agents learn from their work, receive feedback, and coordinate within teams to ensure consistent progress and effective collaboration.