Anthropic Ran 1,000 AI Agents For 8 Days. Yours Still Needs The Same Thing Explained Again.

Anthropic Ran 1,000 AI Agents For 8 Days. Yours Still Needs The Same Thing Explained Again.

More

Summary

AI agents are now reported to work for hundreds of hours, yet many users still find themselves re-explaining the same instructions and fixing the same mistakes. This video from Nate B Jones looks at how to set up long-running agent work so it stays aligned with what you actually want.

He points to recent examples: OpenAI sessions that ran for hundreds of hours fixing problems in its code-management system, Anthropic’s report of more than 1,000 agents working for 8 days on a prototype operating system with Fable 5.2, and Google’s Gemini 4 Carbon working for days on hard engineering problems. He then recounts his own run with Codex, which found 339 sources and wrote over a thousand questions before an agent got stuck repeating work as his goals changed.

The central advice is to pass along evolving context, not a one-time snapshot: decisions made along the way, examples, and instructions that change as your thinking does. He argues for treating projects as team-shaped, with a central repo holding shared data and skills for writing PRDs, proposals and decks, rather than relying on personal “mini-me” setups. That approach also supports human review and accountability for the final deliverable, so errors are caught before they reach a customer.


📺 Source: AI News & Strategy Daily | Nate B Jones · Published November 20, 2026
🏷️ Format: Opinion Editorial

1 Item

Channels

1 Item

People