Abstract

Engineering organizations invest heavily in their postmortems and incident reviews, yet repeated failures persist. We introduce RELAY, a discrete-time agent-based simulation modeling organizational learning from software incidents through a four-stage absorptive-capacity pipeline across an organization over a one-year horizon. We use the simulator to investigate which factors drive incident reduction under controlled conditions rather than to forecast absolute incident rates. We test four organizational levers: sharing scope, deployment velocity, prevention strength, and network topology across roughly 12,000 simulation runs covering our core hypotheses, ablations, and sensitivity sweeps. Sharing scope takes four values (NONE, LOCAL, NEIGHBOR, GLOBAL) We measure the total incidents per organization-year, the performance ordering NONE > LOCAL > NEIGHBOR > GLOBAL holds consistently across as incident counts fall as sharing widens, achieving a 45% reduction from NONE to GLOBAL with a directional minimum pairwise Cohen's d of 3.91. These results demonstrate that sharing scope acts as a fundamental structural gate on organization-wide learning, whereas prevention strength functions merely as an amplifier of knowledge already in motion. Furthermore, broader postmortem sharing buffers against high-velocity deployment risks, and dense network connectivity outperforms centralized hubs. Our replication code is available at https://github.com/pined1/cs-ms-thesis.

Degree

MS

College and Department

Computational, Mathematical, and Physical Sciences; Computer Science

Rights

https://lib.byu.edu/about/copyright/

Date Submitted

2026-08-27

Document Type

Thesis

Keywords

agent-based modeling, organizational learning, software incident management, absorptive capacity, postmortem analysis, knowledge sharing, network topology, DevOps reliability

Language

english

Share

COinS