The reflex in production agent orchestration is to escalate: a cheap model stalls on a hard step, so you hand its trajectory to a stronger one and hope the smarter model digs you out. The Handoff Tax paper measures what that reflex actually buys, and it’s a lot less than the routing diagram promises.
Pairing low-capability and high-capability models from the Claude and GPT families on long-running coding tasks, the authors find that full-trajectory escalation recovers less than half of the quality gap between the weak and strong model — while charging a real cost premium. That penalty is the handoff tax. The strong model doesn’t spend its budget solving your problem; it spends it reconciling another model’s half-finished reasoning. Downshifting, on the other hand — dropping to the cheaper model once the hard reasoning is done — lands on a genuinely favorable cost-quality point.
The detail I’ll actually use is that the best interface reverses with direction. When escalating, reducing the weak model’s inherited trajectory improves the outcome — the strong model does better starting closer to a clean slate than inheriting a muddled one. When downshifting, removing the strong model’s trajectory hurts. So a router that treats “pass the full context along” as the universal handoff rule is wrong in one of the two directions no matter which rule it picks.
For anyone running a tiered agent stack, this argues for asymmetric handoffs: escalate by resetting context and preserving only repo state, downshift by carrying context forward. The HF paper page has the full ablation across handoff timing and interface.
If escalation mostly buys you a bigger bill and a foreign trajectory to clean up, is the stronger model even the right lever — or should a stuck agent re-plan from state before anyone reaches for a pricier model?