A Deterministic Control Plane for LLM Coding Agents

A study of more than 10,000 public GitHub repositories finds that the configuration files steering large language model coding agents are largely unmanaged, with 10.1% of tracked paths being exact duplicates across independent projects and fewer than 1% declaring permission boundaries, according to research published on arXiv [1]. The paper, authored by Padmaraj Madatha, proposes a deterministic control plane called Rel(AI)Build to address what it describes as a governance gap in how LLM coding harnesses operate [1]. The prevalence study examined 6,145 agent configuration files and found that configurations propagate as undeclared shared components: 75.5% of clone pairs crossed organizational boundaries, and 58% of agent configurations were single-commit, meaning they were created once and rarely revised [1]. By comparison, the study found that 33% of GitHub Actions workflows declare permission boundaries, while agent configurations did so in less than 1% of cases, yielding only 31 true positives [1]. The proposed Rel(AI)Build control plane sits above the coding harness and treats agent definitions as a managed supply chain. It uses SHA-256 content addressing, HMAC-stamped lockfiles, and hash-chained audit logs to track configuration provenance [1]. The system enforces tiered permissions and attack-derived blocklists before an LLM is invoked, gates feature work through a phase state machine with requirement-to-file-to-test traceability, and compiles a single canonical definition to seven integrated development environment targets [1]. Prompt drift is detected via Jaccard similarity [1]. Git, the distributed version control system that underpins GitHub, was originally created by Linus Torvalds for Linux kernel development and is now used by nearly 95% of developers as their primary version control system [5]. The research builds on this ecosystem by analyzing repositories hosted on GitHub, which is one of several software forges offering Git repository services alongside SourceForge, Bitbucket, and GitLab [5]. The broader context for the work sits within the field of artificial intelligence, which was founded as an academic discipline in 1956 and has experienced multiple cycles of optimism and funding winters [4]. Since the 2020s, generative AI has become widely available, and the AI boom has coincided with advances in large language models such as Meta's Llama family, which ranges from 1 billion to 2 trillion parameters [4][6]. Reinforcement learning, one of the three basic machine learning paradigms alongside supervised and unsupervised learning, trains agents through interactions with an environment to maximize a reward signal, a concept related to the agent-based systems the paper examines [3]. Conformance tests on injected violations confirmed that each mechanism in Rel(AI)Build enforces its stated invariant, though the paper notes that developer outcomes remain future work [1]. The research argues that governance of this layer must be deterministic and tool-agnostic, not delegated to further LLM orchestration [1].