Testing Interchangeability in LLM Agent Teams

Summary

Tests whether an agent occupying a structural role is substitutable, forming eight independent teams from one base model, letting each agent keep a private notebook over ten formation episodes, then swapping role-matched agents against a placebo that reproduces roster disruption without changing the occupant. Swaps cost little in task score but raise communication per unit of progress by 16 to 63 percent, and the swap penalty moves together with how far independently formed teams drift apart. Relevant because it shows structure has effects invisible to accuracy metrics but visible in coordination and process, which is precisely the paper's argument for measuring epistemic quality rather than benchmark accuracy.

Contribution

Queued — not yet read.

Method

Queued — not yet read.

Relevance to RISE

Surfaced by the 2026-09-07 literature sweep as a medium-priority candidate for the Architecture as Epistemology (ISR) paper (angle: MAS topology / structure vs. aggregation rule). Verified against: Opened https://arxiv.org/abs/2609.05279 and read the full abstract; arXiv API confirms five authors, published 2026-09-04, cs.AI.

Critique / open questions

Queued — not yet assessed.

Key quotes

Production multi-agent systems replace agents constantly, on the assumption that an agent filling a role is interchangeable with any other agent that can do the job. (abstract)

We test that assumption. (abstract)