Opus 5.5 Background Compaction: Agents That Never Stall

Anthropic shipped Claude Opus 5.5 on September 22, and most of the launch-day chatter has been about the price cut ($4 / $20 per million tokens, down from $5 / $25) and the four breaking changes. Buried in the feature list is something that matters more if you run agents for hours at a time: compaction on demand, behind the compact-2026-09-04 beta header. You send one request with a compaction parameter, and instead of a reply you get back a single signed block that summarizes the whole conversation. You then put that block first in messages and carry on.

The part worth building around is that the summary request doesn't have to block anything. You can fire it off in a background thread, keep talking to Claude on the full history while it runs, and swap the block in once it lands. The user never sees a "compacting context..." spinner. This guide builds that as a small, tested Python class, then wires it into a support-triage agent with a tool loop.