Claude Agent SDK gets long-running harness

This is less a new abstraction than a reminder that long-running agent scaffolding has to fit the task and the model’s limits.

• –Anthropic’s setup is a two-part initializer/coding loop, which is easier to reason about than heavier orchestration.
• –The post argues that compaction helps, but it does not solve the harder problem of preserving project state across fresh sessions.
• –The practical pattern is to leave durable artifacts for the next run so the agent can recover context without guessing.
• –For builders, the takeaway is to add structure where agents fail, then trim it back only when the workflow can still hold together.