long-running-agent

The Long-Running Agent framework provides a robust architecture for AI agents tasked with complex, multi-session projects. It solves the common 'one-shot' limitation of language models by implementing a persistent state management system that bridges context gaps between discrete developer sessions. This framework is ideal for developers building autonomous coding agents that need to handle large-scale codebases, incremental feature development, and rigorous testing cycles over extended periods.

• Dual-agent pattern consisting of an Initializer Agent for environment setup and a Coding Agent for incremental progress.

• Standardized session startup protocols using init.sh, feature_list.json, and claude-progress.txt to ensure context continuity.

• Strict state management via progress tracking, which includes logging session timestamps, git commit history, and feature-specific pass/fail verification.

• Automated environment setup through initialization scripts that handle dependency installation and development server orchestration.

• Incremental development methodology that enforces one-feature-at-a-time execution with end-to-end testing requirements.

• Ensure the Initializer Agent is run exactly once to configure the base directory and create essential tracking files.

• Maintain the integrity of feature_list.json by never modifying test descriptions and only updating the status field after successful end-to-end verification.

• Use the provided progress template to summarize changes, document current state, and define clear next steps for the subsequent agent session.

• Integrate browser automation tools like Puppeteer MCP for real-world end-to-end verification rather than relying solely on internal unit tests.

• Always commit progress to git at the end of every session to maintain a resumable, clean state for the next agent instance.

• When adapting to non-coding domains, focus on task decomposition, progress persistence, and verifying work against defined constraints.