An Oxford spin-out building an autonomous AI agent that generates and maintains Java unit tests - using reinforcement learning, not just a chatbot.
Most of the noise around artificial intelligence and software has been about generation - machines that write the code. Diffblue, a small company that grew out of a University of Oxford research lab, went the other way. Its product, Diffblue Cover, writes the tests that prove the code actually works.
Unit tests are the small, repetitive checks that confirm each piece of a program behaves as intended. Developers know they matter and routinely skip them because they are tedious to write and thankless to maintain. Diffblue's bet, placed in 2016, was that this exact drudgery is where automation earns its keep.
Rather than lean solely on a large language model that guesses at code from a prompt, Diffblue built its agent on reinforcement learning - the same broad technique that taught machines to master board games. Pointed at a Java codebase, Cover explores the code, generates candidate tests, runs them, and keeps the ones that compile and pass. Then it maintains them as the software changes.
The customers are not hobbyists. Diffblue says four of the ten largest U.S. banks use Cover, alongside insurers, technology firms and government teams that run large, long-lived Java systems. For that audience, a test that reliably passes is worth more than a clever suggestion that does not.
Figures drawn from company statements and press reports. The 20x and 250x claims are Diffblue's own benchmarks; treat them as vendor figures.
Diffblue Cover reads Java source, writes human-readable JUnit tests, runs them, and updates them as code evolves - working without a developer prompting each step.
Banks, insurers, technology and government teams with large, business-critical codebases. Names publicly linked to Diffblue include Goldman Sachs, Citi, JP Morgan, S&P Global and AWS.
Unit tests are skipped because they are slow to write. Low coverage makes modernizing or refactoring legacy Java risky. Cover fills the gap so teams can change code with a safety net.
Cover uses RL to generate tests that verifiably pass, and - in its 2025 platform - fuses in customer-approved LLMs where they add coverage and quality.
Copilot-style tools suggest snippets a human must accept. Diffblue's agent delivers complete, compiling tests and maintains them - the difference between a demo and production.
A free Community Edition drives bottom-up adoption; Developer and enterprise editions are the commercial B2B business.
"Diffblue Cover is described as writing unit tests up to 250x faster than a human developer - and maintaining them as the code changes."
Diffblue frames itself against LLM coding assistants. Its 2025 benchmark puts the productivity gap for unit testing at roughly 20x on an annual basis, arguing autonomous agents run continuously without a human in the loop. The bars below illustrate that claimed relationship - a vendor figure, shown for context.
Alternatives buyers weigh include GitHub Copilot, Anthropic's Claude Code, Qodo, Amazon Q/CodeWhisperer and Tabnine, plus traditional coverage tooling.
The autonomous agent that writes, runs and maintains Java unit tests using reinforcement learning. The core enterprise product.
Free IntelliJ plugin bringing AI-generated Java tests to individual developers, for open source and commercial use.
A paid edition for individuals and small teams, extending autonomous test generation beyond the free tier.
Lets developers review, accept and refine AI-generated tests inside their workflow - supporting hybrid human/AI testing.
Analyzes existing tests to reuse helper methods and fixtures, generating idiomatic new coverage that fits a team's conventions.
Auto-identifies and prioritizes coverage-blocking issues and produces a turnkey improvement plan - said to lift coverage 50% in an hour internally.
Daniel Kroening and Peter Schrammel spin the company out of the University of Oxford to apply AI to software development.
Led by Goldman Sachs and Oxford Sciences Innovation, funding the AI-for-code platform.
The company commercializes Cover, its automated Java unit-testing product.
Billed as the first AI-powered automated Java unit-testing tool free for developers.
$7M (IP Group) in January and $8M (AlbionVC) in November expand the platform.
Toffer Winslow appointed CEO amid a reported 3x growth period; Developer Edition launches.
Test Review plus a next-gen release and a benchmark claiming a 20x productivity edge over AI coding assistants.
Total venture funding reported by the company at roughly $32M+. Aggregator estimates vary. Figures are approximate.
Enterprise-software veteran with CEO, GM, CRO and CMO roles at firms including Dynatrace, RSA Security, StackState and Tamr. Harvard MBA; leads Diffblue's growth stage.
University of Oxford researcher in formal verification; co-founded Diffblue to bring rigorous, AI-driven automation to software.
Co-founded Diffblue alongside Kroening, drawing on the team's academic roots in program analysis.
Curated searches and channels - opens Diffblue's own video content and demos.
Dossier compiled from public sources · Figures approximate where noted