The company’s latest iteration of Diffblue Cover isn’t just about churning out more tests; it’s about generating smarter tests, faster. This next generation aims to solve the persistent challenge of achieving comprehensive test coverage, a hurdle that often plagues application modernization and regulatory compliance efforts. According to a new benchmark study, Diffblue’s agentic approach, powered by reinforcement learning, leaves LLM-based assistants like Claude Code and GitHub Copilot in the dust.
Diffblue’s new release hinges on three core capabilities, each designed to tackle a specific bottleneck in the testing process. These aren’t mere incremental improvements; they represent a paradigm shift in automated unit test generation.
Test Asset Insights
Imagine an AI that learns from your existing tests, mimicking successful patterns and adapting to your team’s unique style. That’s the promise of Test Asset Insights. Diffblue Cover analyzes pre-existing unit tests, extracting valuable intelligence to generate new tests that seamlessly integrate with your codebase. It reuses helper methods and fixtures, and builds upon domain-specific test infrastructure, resulting in idiomatic, high-quality test coverage.
LLM-Augmented Intelligence
Diffblue isn’t shying away from Large Language Models (LLMs) but is rather embracing them strategically. By fusing customer-approved LLMs with its reinforcement learning engine, Diffblue Cover unlocks a new level of performance. This approach mitigates the security and compliance concerns often associated with introducing new AI models. By leveraging in situ LLMs, Diffblue Cover aims to generate even higher-quality test coverage for complex business logic. An optional standardized approach to connect to LLMs transparently and automate agentic workflows will be provided through a soon-to-be-released MCP server.
Guided Coverage Improvement
Getting initial test coverage is one thing; continuously improving it is another. Guided Coverage Improvement dramatically accelerates the process of expanding test coverage beyond the agent’s initial assessment. This feature helps resolve coverage-blocking issues more efficiently. Internal tests show that this new capability can boost coverage by 50% beyond the initial out-of-box results in just one hour.
“Together, these three new capabilities represent a fundamental shift in how enterprises can deal with under-tested applications,” said Peter Schrammel, CTO and Co-founder of Diffblue. “We’re not just generating more tests faster; we’re orchestrating a complete solution that understands existing test infrastructure, leverages enterprise-approved AI models, and provides intelligent guidance to systematically achieve coverage goals faster and more efficiently than any alternative on the market.”
Diffblue didn’t just pull the 20x productivity figure out of thin air. They commissioned a benchmark study pitting Diffblue Cover against three LLM-based code assistants: GitHub Copilot, Claude Code, and Qodo. The results, according to Diffblue, were decisive.
The study highlighted a critical difference: LLM-based assistants, even in “agent mode,” require constant human intervention – prompting, inspecting, re-prompting, and manually adjusting outputs. This makes them unsuitable for large-scale, unattended test generation. In contrast, Diffblue Cover can be unleashed on a codebase with a single command and left to run for hours, autonomously generating coverage.
The conclusion? Diffblue’s autonomous agent delivers 20 times more code coverage. This, coupled with Diffblue’s fixed-price licensing model, offers cost certainty – a stark contrast to the unpredictable, token-based pricing of LLMs.
Want to see for yourself? Diffblue Cover is available for a self-service trial or a supported proof-of-value evaluation.
Diffblue’s claims are bold, but the potential impact on software development is undeniable. If they can deliver on their promise of 20x productivity gains, they could fundamentally alter the economics of software testing, freeing up developers to focus on innovation rather than tedious test writing. The future of code may well be autonomously tested.
