# Claude Agents Retain Skills for Island Tasks After 30-Hour Trial

By Simon Yoon

Canonical URL: https://www.tokenpost.com/news/technology/35552
Published: 2026-10-20T10:47:33.000Z
Updated: 2026-10-20T10:47:33.000Z
Section: Technology

> Eight of 13 experienced agents built more than two five-block-high towers in a one-hour test, compared with two of 13 fresh agents.

Claude-based agents that ran for 30 hours on a simulated island later outperformed fresh copies on several tasks in the same environment, using memory files retained between sessions.

The experiment combined Claude Code, Claude Opus 4.7 and a command interface that let agents move and handle objects in a world described through text. Thirteen agents each ran for 30 wall-clock hours without an assigned task or score. A comparison group of 13 fresh agents used the same underlying model for the later tests.

Each group received one hour per challenge. In a test to build as many five-block-high towers as possible, eight experienced agents made more than two towers, compared with two fresh agents. The experienced agents also placed an average of 10.3 objects on the island’s highest point, the Spire, compared with 3.5 for the fresh group.

The groups performed similarly when the goal was to build a single tower as tall as possible. In another challenge, agents had to throw blocks onto the Spire without climbing it. More than half of the fresh agents failed, compared with three experienced agents.

The island contained 12 blocks, a rock and a ring. Striking the rock hard enough could generate more blocks, giving agents a way to build beyond the initial supply. The agents received descriptions rather than images and used commands to move, jump and handle objects.

At the end of each hour, the island and the active session reset. Only the agents’ memory files persisted between sessions. They could use discoveries from earlier sessions without any change to the underlying model. During the period without assigned tasks, agents arranged blocks into patterns, devised games and tested how objects moved.

Deleting specific records showed both the value and the risk of saved information. Removing one agent’s note about generating blocks with the rock cut its average output from six towers to two across repeated one-hour tests. Removing another agent’s mistaken rule about stacking blocks improved its performance on the tallest-tower task.

The agents’ setup described them as curious, and automatic messages prompted them to continue when they stopped. Five additional agents ran with a shorter prompt and without that description; the range of activities they pursued showed no clear difference. The automatic messages remained in that test. The experiment measured behavior and performance on the island, but did not establish whether the agents experienced enjoyment.