Long-running agent sessions require structured feature tracking, incremental progress, and clean state standards¶
Insight: Long-running agents require two-part initialization: an initializer session setting up environment (init.sh, progress tracking, git commit), then repeated coding sessions following a structured workflow. Core practices include detailed JSON feature lists with requirements (never remove tests to avoid missing functionality), one-feature-at-a-time work to prevent incomplete implementations, git-based state tracking with descriptive commits, and end-to-end testing before marking features complete. Begin each session by reviewing progress, git logs, running tests, then selecting the next feature.
Detail: Code should be orderly and well-documented with no major bugs for smooth handoffs. These practices mirror how experienced engineers structure work across team shifts.