Anthropic's AI harness represents a significant leap in building long-running AI agents. By overcoming context window limitations, it enables continuous and efficient operations. In a fascinating 24-hour experiment, Claude Code was integrated into this harness to test its real-time application development capabilities.
Project Initialization and Setup
The process kicks off with creating an 'app spec' or Project Requirement Document, laying the groundwork for the project. An initializer agent constructs vital elements like a feature list and a git repository, which houses over 200 test cases essential for project completion. This structured setup ensures that the project starts on a solid foundation.
Continuous Development with Coding Agents
Coding agents play a critical role by reading progress from a 'claude progress' file, implementing features, and performing regression testing. These agents iterate in separate context windows, ensuring that the project is continuously updated and improved. The Claude agent SDK further enriches this process by offering enhanced flexibility, maintaining project directories, and enabling automated web interactions through tools like Puppeteer MCP server.
Results and Future Exploration
At the end of the 24-hour session, the experiment managed to complete 54 sessions with a 54% pass rate on tests. While not fully complete, the complexity achieved was impressive, showcasing functional HTML pages and even theme adjustments. This endeavor encourages further exploration of AI systems with open resources, inviting enthusiasts to experiment and learn structured AI coding workflows.