When people guide AI coding, the tests are better organized

A study of software development workflows found that human–AI teamwork produced stronger test suites than working with AI alone.

Editorial illustration — not from the study.

Researchers compared human-only development, human–AI collaboration, fully automated AI coding and coding carried out autonomously by the MetaGPT X platform. They built test-driven development workflows for each approach and assessed the resulting code, tests and development process.

The human-centred comparison involved 16 IT professionals with experience in test-driven development and Python. Additional repeated executions of the autonomous workflows provided exploratory evidence. The researchers found that the best workflow depended on the goal: autonomous agents were fastest and produced the most functionally correct production code in the study, while collaboration with people produced better-organized tests.

How the workflows compared

The agentic workflow, in which the MetaGPT X platform handled development autonomously, consistently produced the most functionally correct production code according to the baseline test suite and was the fastest approach in the study. It also produced more structurally complex code and lower branch coverage when assessed with the same test suite. The researchers suggest that some uncovered branches reflected implementation choices beyond what the functional specifications required, creating additional untested decision points.

The collaborative workflow, in which developers worked with a conversational language model, produced higher-quality and better-organized test suites, with greater code coverage than the researchers observed in the autonomous approaches. Its production-code correctness was comparable to that of the agentic workflow, but it required more effort because developers often had to revise assertions and coordinate with the AI system. Fully automated, prompt-driven coding was faster than human-assisted approaches but did not match the agentic workflow's production-code correctness or the testing quality of workflows involving people.

Why the choice matters

The findings suggest that there is no single best way to use AI for software development. Autonomous systems may be useful when speed and functionally correct production code are the main priorities, while human involvement remains important when developers need well-organized tests and greater control over how software is checked. Collaboration can balance those aims, but the coordination required between developers and AI systems currently reduces efficiency.

Evidence and caveats

This was an exploratory study. The comparison of human-only and collaborative workflows was a controlled pre-experimental study involving 16 IT professionals from Italy, Denmark and Brazil, all experienced in test-driven development and Python. The fully automated and agentic workflows were assessed through repeated executions on the same development tasks rather than through the same participant-based comparison, so the evidence comes from two different evaluation settings.

The study used structured prompts, Python scripts and the MetaGPT X platform, and the results may not apply to other tasks, AI systems or development environments. The researchers also reported practical difficulties involving an open-ended task, time pressure and platform usability. They describe accurately reproducing development workflows with newer agentic technologies as an open problem.

Programmable hardware speeds matrix calculations up to 6.3 times

Researchers tested a programmable hardware design that imitates selected quantum-computing ideas to run matrix calculations in parallel. In representative benchmarks, the design reached up to 6.3 times the speed and 4.1 times the energy efficiency of conventional matrix co-processors and embedded graphics processors.

2 days ago·13 views

AI & Computing

AI and docking agreed in some drug screens, but not all

Researchers evaluated three AI systems on hundreds of molecules bound to a SARS-CoV-2 protein and in three prospective drug screens. The results suggest that combining AI-based predictions with traditional docking may help prioritize candidates, although no single approach worked best in every test.

4 days ago·50 views

AI & Computing

A hardware monitor detected simulated attacks on car networks in 14 microseconds

Researchers developed a hardware-assisted system intended to detect message injection, spoofing, flooding and impersonation attacks on automotive networks. In tests using simulated traffic and a prototype automotive processor system, it reported an average detection time of 14 microseconds and less than 2 percent CPU overhead.

5 days ago·42 views