A new arXiv paper proposes an architecture for AI-assisted software development that aims to fix structural problems introduced by AI coding agents. The paper, titled "The Spec Growth Engine," argues that while these agents accelerate implementation speed, they also create failure modes that existing spec-driven methods do not fully address.
The visible portion of the abstract highlights one such failure: context explosion, where the agent must reason over an entire codebase or specification, overwhelming its context window. The paper's proposed solution is a "spec-anchored, code-coupled, drift-enforced" architecture, which suggests keeping the spec as the central reference, tying it to the code, and actively preventing drift between them.
The full details of the second failure mode and the mechanics of the engine are not available in the truncated abstract. The paper is listed as a replacement on arXiv, indicating it has been revised, but the provided abstract only gives a partial view.