Four agent frameworks shipped the same fix in one week: context compaction vs. human approvals

I track releases across about 40 agent frameworks, SDKs and personal agents. This week (Oct 12–16), four separate projects shipped work on the same problem: Haystack 3.4.0: an experimental compactor that moves old tool results into a store and leaves a reference in context, instead of summarizing them away. OpenAI Agents SDK 0.24.0: holds off automatic compaction while an approval is pending. Google ADK 2.12.0: require_confirmation for human sign-off before workflow nodes run, plus a callback to filter tool output before it reaches the model. OpenClaw 2026.10.4: manual /compact now lets the active run finish, instead of aborting it and dropping the pending request. My read, from building production agents: this is where long-running agents actually fail. It's not that the model is dumb. It's that the agent forgot what it did 40 steps ago, or killed its own approval request to make room. Compaction is the new retry logic. If you're choosing a framework, ask what it does when context fills up mid-task, and whether a human approval survives that. Questions for people running agents for hours at a time: 1. Do you summarize, offload or truncate old tool output? 2. Have you seen approvals or pending tool calls get lost during compaction? (I wrote this up with links to every release note. Link in the comments if anyone wants it.)