AetherGit: Scalable Source Control for Autonomous AI Agents — A Formal Architecture for Lock-Free, Incrementally Derived Version Control at Agent Scale
Year: 2026
Type: preprint
Abstract: AetherGit is a proposed formal architecture for source control at agent scale, designed to meet the requirements of high-velocity, algorithmically driven development carried out by autonomous AI agents. The architecture is built on four pillars: client-server asymmetry, sparse and virtualized checkouts, lock-free incremental changesets, and a tiered cache topology. The paper gives four formal definitions: (1) the changeset tuple cᵢ = ⟨δᵢ, P(cᵢ), τᵢ, σᵢ⟩; (2) the incremental cost model Cost(cᵢ) = O(|δᵢ|); (3) the recursive derived-data operator Deriv(cᵢ) = ⊕_D(δᵢ, {Deriv(p)}); and (4) the cache-hit probability bound P(Hit) ≥ 1 − (λ_miss/μ_cache)^K · e^(−γ·T_ttl). An illustrative worked example applies the cache-hit bound under stated hypothetical parameters (ρ = 0.6, γ = 0.15). A comparative table positions AetherGit's specified properties against Git's documented design and Scion's publicly disclosed architecture. AetherGit is presented as a theoretical architecture and formalization exercise, developed in direct response to Anthropic's 2026 operational disclosure of its internal Scion system. It does not report an implementation, a benchmark, or empirically measured performance. No implementation has been built and no benchmark has been executed for this manuscript. Worked example status: Section 6 is an illustrative application of the formal bounds under stated, clearly labeled parameters, not a report of measured performance. The paper's contribution is a formal vocabulary and a set of provable structural claims, offered as a reference point against which a future implementation, or other systems serving agentic software-engineering workloads, can be evaluated. Version 1.0.0 — November 2026.
Source: Figshare
Author: Samir Baladi
Language: English
Subfield: Information Systems
Field: Computer Science
Domain: Physical Sciences
Open Access status: green
OpenAlex ID: W7225338604