Anthropic: initializer + incremental coding across context windows
How to keep long-running coding agents coherent: feature lists as JSON, git+progress artifacts, clean handoffs, and end-to-end testing as a human would.
Core problem: discrete sessions with amnesia. Compaction isn’t enough—agents one-shot until context dies, or later sessions declare victory early.
Two-part harness: initializer writes init.sh, progress log, git baseline, and a structured failing feature list; coding sessions pick one feature, leave clean mergeable state, commit with messages, update progress. Strong instructions not to delete tests. Start each session by running fundamentals end-to-end (browser tools) before new work.
Table of failure modes mapped to initializer vs coding behaviors is the reusable artifact. Open question: specialized tester/QA agents vs one general coder.
Why it matters for code review
Incremental, artifact-backed progress is what makes multi-session agent work reviewable. Without clean handoffs and honest feature status, reviewers inherit half-built messes and false “done.”