Scion: Rethinking Source Control for Agentic Software Engineering

Scion is a position paper and proposed empirical evaluation protocol examining the architectural transformation required when source control systems serve parallel, agentic workloads rather than human-paced collaborative development. The paper treats Anthropic's October 2026 disclosure of Scion (an internal source-control system) not as a product announcement, but as a case study motivating a broader architectural question: what structural transformations become necessary when code production shifts from a human-paced, sequential activity to a parallel activity driven by autonomous agents? The analysis is organized around five axes: workload characteristics, repository architecture, data architecture, execution infrastructure, and governance/security. The paper proposes a three-layer design: - Layer A (Architectural Analysis): What source control must do differently under agentic workloads. - Layer B (Structural Analysis): How the five axes interact and constrain each other. - Layer C (Empirical Evaluation): A proposed, unexecuted protocol for future benchmarking. Three critical corrections are built into the analysis as binding constraints: 1. Scion did not eliminate merge conflicts; it shifted them to post-merge/pre-release testing. 2. Scion is not unconditionally lock-free; changeset creation avoids full coordination, but ref updates still use CAS. 3. Derived data is treated as an architectural decision, not a storage optimization. Two open research questions are posed: 1. What is the latency cost of on-demand derived-data computation? 2. What is the scope of "functional decentralization" in practice? All operational figures attributed to Anthropic are explicitly flagged as self-reported, not independently verified. Resources: - Zenodo DOI: 10.5281/zenodo.23934465 - Figshare DOI: 10.6084/m9.figshare.34908594 - PyPI: https://pypi.org/project/rethinking-source-control/ - Website: https://scion-claude.netlify.app - GitHub: https://github.com/gitdeeper15/Scion License: MIT (code) + CC-BY-4.0 (manuscript, figures, documentation). Author: Samir Baladi (ORCID: 0009-0003-8903-0029)