Speedscale Proxymock: Freely testing cloud native apps alongside AI code assistants

We’ll always remember 2025 as the year AI code assistants went big. Copilot, Cursor, Claude, Windsurf, whatever. Developers went from mistrusting these tools, to being expected to turn over much of their coding labor to them.

Even if, according to an extensive Stack Overflow survey, only 3 percent of professional developers say they ‘highly trust’ AI coding tools.

Companies expect these GenAI coding platforms to replace most of the grunt work of development—so much so, that many IT departments have stopped hiring junior developers—at least for now. Which also means, there won’t be enough mid-level developers around these shops in a few years to review the work, and separate the elegant code from AI slop.

AI adoption is rising, but is it enhancing developer productivity?

According to the latest 2025 DORA State of AI-assisted software development report, 90% of the respondents to a study about AI development tools happen to be using them already. More significantly 80% of developers using AI cite a productivity boost.

But how much of a productivity boost? You can’t really measure productivity by lines of code generated, or even the number of releases. An 8-12 percent improvement in feature points delivered would be significant, but it’s a far cry from the 10X gains we were being promised.

Perhaps the problem is how quickly AI can generate code and make changes, faster than the best test automation tools can catch up with, even newer agent-based testing solutions. AI-generated code doesn’t just mean we are adding more code to be tested, it means we are simultaneously destabilizing the environment the code runs in.

Every prompt sets off an API request, which may set off many more requests to other services and sources of information. Which means tokens on more inference, and more API calls to other agents, services, and data services.

If you’re lucky, the coding assistant might recognize a problem in its reasoning and try again, and it might respond in a couple minutes—not long enough to get a coffee, but long enough to sit there and wait, and if the results aren’t what you expected, you can ask it to test itself and wait some more.

Hope it doesn’t get stuck in a loop. Cursor or Claudecode may for instance make a coding mistake, and then repeat that mistake when devising a test, so the test passes anyway. The newly generated code and the test script become technical debt.

It’s a vicious cycle which can cost your project time and money.

Instead, why not give your AI dev partner a ready life-like environment based on real user data, so there’s no guessing about success or failure.

With Proxymock, realistic traffic, API mocks and AI coding sandboxes are free

What happens when you (and your coding assistant) are confident enough to go to production? You need to run enough tests to validate that the application will work within its target environment, with a high degree of confidence.

That’s exactly what Speedscale does with their enterprise platform: capturing and replaying real production traffic, simulating test data, and generating tests in Kubernetes clusters that can scale to thousands or hundreds of thousands of instances, and shut down just as quickly.

If you are bringing AI coding assistants and agents into your shop, these replayable environments will also be a lot cheaper than telling an LLM to burn millions of tokens testing itself to see if the new software will fall down in production.

Still, that might be overkill when you are just working with a copilot and trying to validate the code changes to see if it “hallucinated” any changes, or decided on its own to change a library in some non-deterministic way. You know, good old smoke and regression testing, to make sure fixing the oven doesn’t blow up the toilet.

For these situations, Speedscale introduced Proxymock, a free tool that you can embed in the CI/CD pipeline or git repo to instantly pop up a single test environment that includes mock API endpoints and request/response data pairs that are either synthetically generated or captured from real traffic.

Platform engineering for agents

I’ve written quite a lot about platform engineering practices, which replace the old “dev tools” team, and create self-service shared resources for development environments to encourage code, component, and configuration reuse. These practices aim to simplify the development process by providing developers with self-service access to tools and environments, reducing setup costs and time.

Proxymock also talks to an MCP server—so any AI code assistant or agent can see that Proxymock tools for testing are installed right there in the neighborhood, and grab a mock to check its own work within the context of the application under test.

Of course, if you want to horizontally autoscale those proxymocks into the tens or hundreds of thousands, you can still buy their commercial tool for that.

As it turns out, the old ‘fake it till you make it’ maxim of service virtualization that started taking hold way back in 2007-2010 when I worked at ITKO with Speedscale’s founder Ken Ahrens has become even more effective today.

Thanks to the addition of real user monitoring, synthetic test and test data generation, cloud native portability, Kubernetes scalability, and now AI code assistants and agentic AI frameworks, our testing systems are evolving nearly as quickly as the development platforms they need to test. That’s why it’s been so much fun to follow Speedscale on this journey.

Guest Article by Jason English. Jason is writing this guest article as an advisor to Speedscale. He is a software industry analyst at the firm Intellyx, and former head of marketing for several development and testing software firms.