Why We Built a Compaction System Instead of Handing Off to New Agents
aiclaude-codeopenrigmulti-agent
When a long-running orchestrator hits context limits, the obvious move is to hand off to a fresh agent. We tried that. The startup overhead, lost momentum, and broken continuity made it worse than the problem it solved — for our use case. So we built a compaction system instead: one that writes ground truth to disk, restores from it, and distills session knowledge into a wiki page via an API call to Haiku. Recent research shows neither approach is categorically better — compaction drops behavioral constraints, handoff carries a 'handoff tax' — but for same-model orchestrators that accumulate judgment over time, compaction with external ground truth wins. Here's how the system works and what the literature says.
