AWS Paper Exposes Agent Handoff Tax
The paper measures how switching models mid-task affects coding-agent quality and cost across Claude and GPT families. Full-trajectory escalation recovers less than half the stronger model’s quality advantage while adding substantial cost.
Model routing is not just a capability problem; inherited trajectory can actively handicap the receiving model.
- –Full context is not always helpful: trimming the weaker model’s trajectory improves escalation.
- –Downshifting works better when the stronger model’s trajectory is preserved.
- –SWE-bench results suggest restarting with a stronger model may beat continuing a failed run.
- –Agent frameworks should track handoff quality, recovery rate, latency, and cost separately.
- –The practical lesson is to transfer task state selectively, not blindly copy conversation history.
DISCOVERED
1h ago
2026-08-26
PUBLISHED
2h ago
2026-08-26
RELEVANCE
AUTHOR
omarsar0