Coding & software factories / ALLY AI FRONTIER
DemonstratedWhen AI teams build software together.
Anthropic’s compiler experiment and StrongDM’s software factory demonstrate coordinated AI work. People shape objectives, testing environments and the criteria for a useful result.
Source published: 2026-02-05
Editorially reviewed: 13 Oct 2026
AI illustrationOriginal AI concept artwork. The linked sources show the actual research results and products.
What has been achieved?
In February 2026, Anthropic reported an experiment with sixteen AI agents working in parallel. They developed a C compiler in Rust that could build a Linux kernel under the documented conditions. A researcher designed the working environment and improved its tests. Separately, StrongDM describes a software factory in which specifications and application scenarios guide agents. Together, the examples show a broader workflow: software taking shape through coordinated steps, feedback and repeated assessment, rather than only through individual answers to programming questions.
Why does it matter?
An interesting question for teams is which responsibilities they want to shape deliberately. The people who talk with users, explain priorities and decide what a product should do also shape the agents’ work. Our focus is the quality of that working experience: clear assignments, considered handovers, helpful feedback and time to learn. A software factory provides a reason to discuss those choices. Its usefulness in an organisation should be assessed with the people who develop, review and eventually use the software.
What remains open?
Anthropic explicitly presents its compiler as a capability test and documents technical limits. It is not a complete replacement for an established compiler; parts needed existing tools, and new changes still introduced errors. StrongDM’s report describes that company’s approach. It does not establish a general recommendation to remove human review. Real products need accountable decisions about quality, security and maintainability. These examples demonstrate interesting forms of coordinated agent work, without establishing a universal staffing reduction or guaranteed productivity improvement for other teams.