Historical brief · 2026-11-17

Agents Cannot Hand Over What They Have Not Yet Understood

The authors report in this preprint: Researchers study multi-agent knowledge transmission across generations in 56 interactive worlds, finding that repeated summarization leads to cumulative information loss that standard prompting fails to fix.

Source image · arXiv Artificial Intelligence

The authors report in this preprint: A preprint study investigates how well AI agents transmit learned knowledge to successors across generations using 56 interactive worlds with hidden rules. The authors compare a single agent playing eight consecutive lives, retaining its full history, against a chain of eight agents where each plays one life and inherits only a 300-character note. Across all six tested models, the chain learns less, reaching a performance level after eight generations equivalent to what a single agent achieves after four lives. Controlled experiments decompose this handoff loss into components: changing the agent introduces no loss, single-step compression to 300 characters incurs a small loss, and repeated compression across generations causes a major loss. When each note is generated from the complete cumulative record of all previous lives rather than iteratively from the prior note alone, 70% to 82% of the lost performance is recovered. The authors note that this degradation persists across model scales and is not mitigated by instructing the writer to record evidence, revise inherited notes, or maintain a structured ledger.

View selection rationale
PreprintStage not established

Importance 77/100

Signal80Structure80Consequence70Confidence75Momentum50
About this brief
Content type
Research summary
Source type
First-party / official source
Source published
Process
AI-assisted discovery, translation, summarisation, classification and scoring; standards and corrections are maintained by the editorial desk.
Verification
The original URL and publication metadata are retained. This brief is not independently fact-checked unless explicitly stated.
Published by
Polyemergence Editorial Desk · PolyEmergence Pte Ltd.
Conflict disclosure
No financial or commercial conflict declared.

Editorial standards → · Corrections →

POLYEMERGENCE / EXPLORE

Keep exploring

Follow the topic, not just the headline.

Explore the Artificial Intelligence archive →