Tracking the Behavioral Trajectories of Adapting Agents

Jonah Leshin, Manish Shah, Ian Timmis · 2026

Read Tracking the Behavioral Trajectories of Adapting Agents by Jonah Leshin, Manish Shah, Ian Timmis (2026) in the Tower of Babel, a free online library of public-domain works preserved in clean, searchable text. No sign-up. The full text is served as a plain-text file at a predictable URL, where text is present: https://antarctic-labs.com/tower-of-babel/arxiv-2606-02536.txt.

Description

Leshin and co-authors measure how agent behavior traits shift as text files change. Modern agents are steered by skill files, memory files, and configuration files. Humans and agents edit those files over time. The edits steer future behavior. The team defines traits as directions in an embedding space.

A linear model learns a trait vector from labeled before-and-after skill diffs. Any new edit is scored by projecting its diff onto that vector. The test trait is a tendency to seek sensitive data. On 68 labeled skill diff pairs, sign accuracy reaches 91.2 percent. Rank correlation reaches 0.82 under leave-one-out testing. The method plugs into an agent-to-agent protocol. One agent can judge another agent's skill updates through a trusted intermediary. The work gives a concrete way to track behavioral drift in agents.

Catalog record

CREATORJonah Leshin, Manish Shah, Ian Timmis
YEAR2026
COLLECTIONPAPERS
FORMATTXT
FILE SIZE20 KB
LICENSEhttp://creativecommons.org/licenses/by/4.0/
RIGHTSCC BY
SOURCEarXiv

Source: https://arxiv.org/abs/2606.02536

Browse the full library · About the Tower of Babel