Raising test coverage at industry scale is difficult: engineers report a median of four minutes to produce a single covered line of code, making coverage targets expensive in large repositories. We present AutoCover, a production system that automatically generates, validates, and proposes tests using large language models. AutoCover supports three complementary interaction modes: a CLI for local scripting, a Headless mode for generating tests at scale across repository shards and creating merge requests, and an IDE mode for human-in-the-loop generation that captures developer intent and enables rapid fixes. Together, these modes support both legacy code in repositories and newly developed code on developer machines.

AutoCover is implemented as a modular, agentic pipeline built on LangGraph, with subgraphs for preparation, generation, execution, and validation or repair. Preparation includes scaffolding tests and running initial coverage, while generation uses multi-shot prompting informed by prior failures. Repository adapters provide language- and repository-specific actions such as import splicing, coverage commands, and coding conventions, and a code-context retriever supplies only relevant symbols to the LLM to stay within context limits. To reduce low-quality outputs, AutoCover combines intent-aware generation with validation gates such as coverage deltas and mutation or branch checks where available, along with flakiness defenses like multi-run CI simulation when compute permits.

This paper makes three contributions. First, it describes AutoCover’s end-to-end architecture and evolution, including design considerations around user experience, test quality, and cost. Second, it details interaction patterns across CLI, Headless, and IDE modes, and shows how intent collection and human-in-the-loop repair improve acceptance. Third, it reports both intrinsic and extrinsic evaluation results. AutoCover now generates about 11% of all new tests that are reviewed and added to CompanyX’s codebase.