Published benchmark research from generative AI coding pioneer quantifies compelling advantages of new solution’s end-to-end process automation over stand-alone coding agents

Diffblue today announced the general availability of the Diffblue Testing Agent, an autonomous regression test generator that works with an enterprise’s existing AI coding platform — GitHub Copilot, Claude Code, and others — to generate verified unit tests across entire codebases without developer intervention. In benchmark testing across eight real-world Java projects, the Diffblue Testing Agent automatically delivered 81% average line coverage compared to just 32% achieved by a senior developer iterating with an AI coding agent alone – a 2.5x coverage advantage and major developer productivity upgrade.

Enterprise adoption of AI coding tools has accelerated dramatically, but the challenge of automatically generating large volumes of trustworthy code remains a persistent unsolved problem. This pain is particularly acute when attempting to create comprehensive regression unit test coverage – a foundational requirement for any production-ready code and application modernization initiatives. Until today, developers have struggled to coax even modest levels of test coverage from coding agents, often resorting to iterative prompting that delivers rapidly diminishing returns while incurring substantial token costs.

The Diffblue Testing Agent addresses this problem by delivering an orchestration and verification layer on top of the AI coding platforms enterprises have already deployed. Rather than replacing an organization’s existing tools, the Diffblue Testing Agent works alongside them delegating method- and class-level test creation to the underlying AI coding agent while orchestrating a comprehensive process that includes coverage analysis, build system configuration, test plan creation, parallelized test generation, output verification, project clean-up, and pull request preparation. The agent autonomously scopes the codebase, generates tests, verifies that every test compiles and passes, and rolls back any that don’t — across hundreds or thousands of classes in a single run, without developer intervention. The autonomous workflows implemented by the Diffblue Testing Agent are built on decades of software verification expertise and research originating from the University of Oxford.

The impact of this new approach is compelling. Newly published benchmark research quantifies the advantages delivered by the Diffblue Testing Agent by measuring its performance when used to generate test coverage for 8 real-world code repositories compared to the experience of a developer using just Claude Code to generate test coverage. Diffblue’s agent was deployed with a single prompt and ran autonomously on each project generating an average of 81% line coverage and 61% mutation coverage – both of which exceed typical enterprise standards. In contrast, an experienced, AI-savvy developer was only able to deliver an average of 32% line coverage and 24% mutation coverage after two hours of toil iterating with Claude Code to keep it on task and verify its output – a process that is painful and an outcome that is far short of target. The full benchmark study, including methodology, project-level results, and codebase details, is available at diffblue.com/benchmarks.

“The industry has spent two years proving that AI can write code. The question now is whether AI can do actual engineering work — reliably, at scale, without constant human supervision,” said Dr. Peter Schrammel, Diffblue co-founder and CTO. “Our benchmark data shows that the developer effort for driving even the best AI coding agents reaches unaffordable levels quickly. For writing regression tests at scale, diminishing returns make progress beyond 50% coverage difficult, whereas the Diffblue Testing Agent achieves 80%+ autonomously. That gap is the difference between an AI experiment and an AI-enabled engineering workforce. We built the orchestration layer that closes it.”

Diffblue’s Testing Agent is available immediately for enterprise evaluation. The initial release includes autonomous regression unit test generation for Java and Python. The product integrates with GitHub Copilot and Claude Code, with additional AI coding platform support planned based on customer demand. In the coming quarters, the company will extend its Diffblue Agents platform to additional software quality domains, including test quality assessment, code review automation, large-scale refactoring support, and requirements-driven test generation. Development teams interested in evaluating Diffblue Agents can contact sales@diffblue.com or visit diffblue.com/agents.

Diffblue makes AI coding agents reliable at enterprise scale. The company’s Diffblue Agents platform orchestrates autonomous software quality workflows on top of the AI coding platforms enterprises already use, starting with automated regression test generation for Java and Python codebases. Diffblue enables innovative development teams everywhere to innovate faster, modernize with confidence, and improve software quality. Founded by award-winning researchers from the University of Oxford, Diffblue is backed by leading investors including IP Group, Albion, Parkwalk, and Citi.

View source version on businesswire.com: https://www.businesswire.com/news/home/20260324227278/en/

Media Contact:
Cassandra Locke
cassandra.locke@diffblue.com

Logo

Tags

More Like This

Diffblue’s Latest Innovations in Unit Test Generation Deliver 20x Productivity Advantage Versus AI Coding Assistants

New Diffblue Integration Delivers Enterprise-grade Fully Autonomous AI Java Unit Testing for Easy Continuous Integration

Diffblue Receives £1 Million Grant from Innovate UK to Advance AI-Driven Software Engineering

Diffblue Announces Citi as a Strategic Investor

BrowserStack Launches Suite of AI Agents to Redefine Software Quality at Scale

LambdaTest Unveils the World's First Platform to Test AI Agents: Introducing Agent-to-Agent Testing

BrowserStack's AI Agent Cuts Test Case Creation Time by 90%

BrowserStack Bridges Developer-QA Productivity Gap with Test Failure Analysis Agent