Skills
For agents that span multiple context windows, replace compaction with durable structured artifacts (feature-list JSON + git + progress files)
Anthropic's long-running-agent harness work found that context compaction alone fails because it doesn't hand clean instructions to the next session; instead an 'initializer' agent writes durable, queryable records (a JSON feature list, git history, a progress file) that a 'coding' agent parses fresh each session. For builders running overnight or multi-day agents, the takeaway is to stop trying to preserve context and start externalizing state into files the next run can re-read deterministically.
↳ Follow the thread