According to Ethan Mollick on X, Claude with Computer Use autonomously downloaded the NetHack codebase, read documentation, and began implementing a new horror-inspired creature by modifying source files until hitting rate limits, demonstrating concrete agentic capabilities for software development workflows (as reported by Ethan Mollick’s X post and thread). According to Mollick’s post, the model executed multi-step tool use including repository fetch, file inspection, and targeted code edits, highlighting near-term applications in rapid prototyping and legacy code maintenance for game development and enterprise software. As reported by Ethan Mollick, the run-by-run trace suggests viable business use cases such as automated feature insertion, refactoring, and test generation under human supervision, with constraints around API rate limits and oversight requirements.

According to Ethan Mollick on X (Twitter), OpenAI's Codex autonomously downloaded NetHack, modified game items to increase player power, and produced a working Windows .exe, overcoming environment and build issues that previously stymied older AI tools. As reported by Mollick’s post, this showcases practical code synthesis, dependency management, and build orchestration—key capabilities for AI software agents. For businesses, this indicates near-term opportunities to automate legacy app refactors, rapid prototyping, and modding workflows; according to Mollick, the successful artifact delivery (.exe) is evidence of reliable multi-step tool use that can reduce developer cycle time and QA overhead in controlled pipelines.