The AI Velocity Paradox: Mitigating Specification Drift in Autonomous Coding Agents via Deterministic Software Intent Graphs

Autonomous artificial intelligence coding agents (e.g., Cursor Composer, Claude Code, OpenAI Codex, OpenCode) generate implementation code and unit test suites at unprecedented speeds. However, this velocity introduces an unaddressed structural defect: silent specification drift. When stochastic large language models (LLMs) author both business logic and test assertions without an external, verifiable contract, they commit circular testing—synthesizing tautological unit tests that validate hallucinated assumptions rather than human intent. Furthermore, common heuristic context injection techniques like Vector Retrieval-Augmented Generation (RAG) provide semantic proximity rather than normative constraint. In this paper, we formalize the Software Intent Graph (SIG): a Git-native, bidirectional, deterministic verification harness that bounds probabilistic coding agents. By marrying formal Requirements-as-Code (StrictDoc) with incremental, Tree-sitter-based Abstract Syntax Tree (AST) code extraction (Tracey 2.x), we map human specifications to implementation functions (r[impl]) and verification tests (r[verify]). We expose this intent topology directly to autonomous agents via the Model Context Protocol (MCP) and enforce non-negotiable verification gates in CI/CD. We evaluate our framework across an empirical benchmark of 40 enterprise software modification tasks executed by autonomous agents. Key findings demonstrate: Unconstrained Baseline: Agents exhibit a 42.5% silent specification drift rate and produce tautological tests in 35.0% of cases, despite achieving 94.2% statement coverage. SIG-Governed Pipeline: Reduces silent specification drift to 0% at pull-request merge time. Traceability & Verification: Detects 100% of broken or stale references with a sub-50ms daemon query latency overhead. Companion Monograph: The extended 20-chapter operational curriculum, full test harnesses, and production architectural blueprints are documented in the companion book: Governing AI Coding Agents: Spec-Driven Development, Intent Graphs, and Automated Traceability (Leanpub).