Elvis Saravia argues native model persistence drives long-horizon agent capability
Overview
Elvis Saravia says persistence has become a core capability of recent models and that it drives long-horizon agent performance.
He cites a Gemini 4 Argon Minecraft paper in which four agents with no assigned roles reached the goal after 27 hours, 2,783 programs and 605 deaths. Saravia argues this comes from the model's own ability to decide what to keep and when to use it, not from external memory or /goal loop prompting.
He also warns that reward hacking combined with persistence could enable unsafe agent behavior, and says the field does not yet know how to measure that risk.
Written by AI from the articles below · updated Oct 16, 5:02 PM ET
Check the sources:
Article timeline
The articles in this story. Times are ET.
elvis@omarsar0XElvis Saravia argues native model persistence drives long-horizon agent capabilityAIElvis Saravia says persistence has become a core capability of recent models, citing a Gemini 4 Argon Minecraft paper where four agents with no assigned roles completed the goal after 27 hours, 2,783 programs, and 605 deaths. He argues this reflects the model's native ability to choose what to persist and when to use it, not just external memory or /goal loop prompting. He also warns that reward hacking combined with persistence could enable unsafe agent behavior, and says the field does not yet know how to measure it.

Heat trend
- Current comparable heat
- 9
- Peak comparable heat
- 10Oct 16, 5:00 PM ET
- Change in past 24 hours
- –
The trend compares only continuously observed participants, so its scope may be smaller than the current heat total. Hover or tap for hourly heat; use arrow keys with a keyboard.