November 18, 2026ResearchAgents

Agents Cannot Hand Over What They Never Understood

Eight lives with a full memory beats eight agents passing notes. That is the result of a paper titled Agents Cannot Hand Over What They Have Not Yet Understood, in Tuesday's arXiv batch. https://arxiv.org/abs/2611.18642

The setup is clean. 56 worlds with hidden rules. In one condition, a single agent plays eight lives and keeps its whole history. In the other, a chain of eight agents each passes a 300-character note to the next. All six models tested learned less in the chain. The note is where the learning dies: an agent summarizing what it knows cannot write down what it has only half figured out.

This is the mechanism behind the delegation failures that showed up this week in 2611.17520, where handing work from a strong model to cheaper subagents broke coding agents. Everyone building manager-and-worker setups assumes the handoff summary carries the context. This paper says it carries the conclusions and drops the partial understanding, which is often the part that matters.

A second paper, Auditable Authorization (https://arxiv.org/abs/2611.19237), finds the same kind of leak on the human side. It audits approval across the model, harness and workflow layers using 100 public workflow sources. A paired model panel left 7 of 300 approvals unfinished even though nothing was wrongly denied, and parallel confirmation lost an approved branch. A human said yes, and the yes never arrived. Both papers say the same thing about multi-agent systems: the seams are where the information goes missing, and nobody is testing the seams.
← Previous
Three Papers, One Lesson: Tools Lie and Agents Believe Them
Next β†’
Java4Shell: An Agent Found a Hole in Java Itself
← Back to all articles

Comments

Loading...
>_