Uber developed AutoCover, an LLM-based multi-agent system for generating software tests, and gradually deployed it to all of its engineers between September 2024 and September 2025. In a paper presented at ICSE 2026, its authors report that AutoCover generates about 11% of new tests reviewed and added to Uber’s codebase. This is the authors’ reported share of reviewed and added tests, not a measured reduction in bugs. As of September 2025, the authors report monthly volumes of hundreds of thousands of tests in the IDE, tens of thousands in headless runs and thousands through the CLI.
The system uses LangGraph to orchestrate five specialized agents: a preparer discovers targets and scaffolds tests; a generator produces test cases; an executor runs them; a validator checks conventions, coverage and available mutation tests; and a fixer repairs failures. Repository-specific adapters connect the agents to build rules, test commands and coding conventions. A context retriever selects relevant code symbols for the language model, and deterministic rollback limits unsuccessful edits.
AutoCover offers three modes: a CLI for on-demand local testing, a headless mode that generates tests across repository shards and creates merge requests for owning teams to review, and an IDE mode that provides interactive generation and repair. The paper reports viable-test generation rates of approximately 20% for Java, 40% for Go and 80% for Python, and says users explicitly accepted about 44% of viable IDE tests. These rates describe the authors’ evaluation, not a guarantee for other environments.
Details
- City
- Seattle
- Organization
- Uber
- Continent
- North America
- Country
- United States
- Category
- Ground Transportation