Anthropic details long-running agent harness

This is Anthropic turning agent reliability into a systems problem, not a prompting problem.

• –The initializer-agent pattern front-loads structure so later sessions inherit a usable workspace instead of guessing at project state
• –The progress file plus git history gives the next agent a durable memory layer, which is exactly what long-running workflows have been missing
• –Forcing one-feature-at-a-time execution reduces the classic failure mode where agents try to one-shot an entire app and then strand the repo
• –The testing guidance matters as much as the scaffolding: end-to-end verification is what separates “looks done” from actually shippable
• –The pattern is tool-agnostic enough that teams can adapt it beyond Claude, which makes it more interesting as infrastructure than as a one-off Anthropic demo