Anthropic has showcased significant advancements in autonomous AI development through an innovative experiment involving sixteen AI agents that independently constructed a C compiler. Conducted using the Claude Opus 4.6 model, the experiment illustrated both the progress and limitations of such technologies at a time when several AI providers, including Anthropic and OpenAI, are prioritizing agentic systems.
The ambitious experiment involved the AI agents building a C compiler in Rust from the ground up. Following the establishment of the project goals, human supervisors largely removed themselves from the process, allowing the agents to collaborate in parallel on a shared Git repository, devoid of a central overseeing agent or orchestration.
This independence was made possible by Anthropic’s technical infrastructure, where each AI agent operated within its own Docker container and repeatedly executed tasks in an infinite loop. Coordination among the agents relied on simple lock files in the repository, ensuring that they did not interfere with each other’s work.
Over nearly two weeks, the project encompassed about two thousand Claude Code sessions, processing approximately two billion input tokens and generating around 140 million output tokens, which amounted to nearly $20,000 in API costs. The resulting compiler, consisting of roughly 100,000 lines of code, demonstrated its capabilities by successfully compiling a bootable Linux 6.9 kernel for x86, ARM, and RISC-V architectures. Other major software projects, including PostgreSQL, SQLite, Redis, FFmpeg, and QEMU, were also compiled with success. Notably, the compiler achieved a success rate of approximately 99 percent on the GCC torture test suite and even compiled and ran the classic game Doom.
However, external commentators have raised questions about the degree of autonomy exhibited by the AI agents. Although they were capable of writing code independently, significant human effort went into the preliminary design of test harnesses, continuous integration pipelines, and feedback mechanisms that accounted for the limitations of language models. Ars Technica noted that the majority of the work revolved around this preparatory phase rather than programming itself.
Anthropic asserts that the compiler was developed without direct external influences, stating that the AI agents lacked internet access during the coding process and relied solely on the Rust standard library. This has led the company to refer to the project as a clean-room implementation. Nonetheless, this characterization is contentious; the underlying language model was pre-trained on substantial amounts of publicly available source code, likely including existing C compilers and associated tools, which deviates from the classic definition of a clean room in software development.
As the project progressed and the code base expanded toward 100,000 lines, it became evident that new bug fixes and extensions frequently disrupted existing functionalities. This phenomenon, familiar in human-generated codebases, also appeared in the AI agents’ work, suggesting a practical scale limit for autonomous software development with current AI models.
The complete source code from the experiment is publicly accessible, with Anthropic framing the project as a research initiative. The findings demonstrate the potential of contemporary AI agents while simultaneously highlighting the challenges of large-scale autonomous software development. As the industry continues to evolve, these insights could shape future innovations and applications in AI-driven solutions.
For more information on Anthropic, visit their official site at anthropic.com.
See also
Germany”s National Team Prepares for World Cup Qualifiers with Disco Atmosphere
95% of AI Projects Fail in Companies According to MIT
AI in Food & Beverages Market to Surge from $11.08B to $263.80B by 2032
Satya Nadella Supports OpenAI’s $100B Revenue Goal, Highlights AI Funding Needs
