The 71-Cent Handoff

A paper called The Handoff Tax, out of AWS’s Agentic AI group on 25 August, asks a narrow question at unusual scale: when a coding agent switches models partway through a task, what happens to the model that inherits a trajectory it did not write? Two model families, seven switch points, four handoff interfaces, 58 configurations per family — 58,000 runs and 36 billion tokens on SWE-bench Verified. The abstract is diplomatic, saying full-trajectory escalation “recovers less than half of the LC-to-HC quality gap while incurring a substantial cost premium.” The table is not. For the Claude pair: the cheap model alone solves 60.7 percent at $0.40 a task; the strong model alone, 79.2 percent at $0.72. Escalating mid-run by passing the full trajectory forward — the default, the thing a /model command does — lands at 69.2 percent for $1.61.

The row that matters is one the abstract never mentions. The authors added a control that throws the cheap model’s work away, restarts the strong model from the original task statement, and still bills you for the discarded work: 79.2 percent at $0.90. A second control lets the cheap model run to completion first, then restarts: 79.2 percent at $1.12. Both beat mid-run escalation on both axes. Escalation costs 79 percent more than abort-and-restart and solves ten points fewer. That is not a cost-quality tradeoff, it is a dominated cell — and the dominance survives paying for the throwaway. The sunk cost is not sunk; it is negative. You are spending roughly seventy-one cents a task for the privilege of letting the strong model read the weak model’s notes. The uncomfortable part is that essentially every cost-conscious agent design shipping today is a cascade — try cheap, escalate when it struggles — and every one of them is a bet that a routing error can be cheaply corrected in flight.

The mechanism is stranger than the price. Consider one interface, traj-drop: discard the conversation, keep the edits already made to the working tree. Going up the capability ladder it is the best interface available, lifting quality recovery from 47 to 64 percent for Claude and 36 to 84 for GPT. Going down the ladder it is the worst, scoring 28 percent where simply passing the raw trajectory scores 50. Same knob, same bytes, and the optimum inverts with direction. The authors call this directional duality. The way I would put it is that inherited context carries a sign, fixed by who wrote it relative to who is reading it: a strong model’s reasoning is scaffolding for a weak one, and a weak model’s reasoning is an anchor on a strong one. The economics are equally direction- and vendor-specific. Raw escalation costs 2.2x more than starting strong for the Claude pair ($1.61 against $0.72) but less for the GPT pair ($0.36 against $0.47), purely because Sol runs about 8x Luna while Opus runs about 2x Haiku. “Should I escalate” has no model-independent answer; it flips on the price ladder.

Follow that out and routing loses most of its option value. You cannot buy information by trying the cheap model first, because the trace you generate while learning is itself a liability on the balance sheet. The decision has to be right before the first token, which quietly converts routing from a model-selection problem into a difficulty-prediction problem on a prompt nobody has seen yet. Which also adjudicates the week’s other argument. Small Models Have Arrived (702 points on HN) reports routine agent work falling from about a dollar a run to about a dime; Terminal-Bench-Science, released the same week, scores the small model at 3.3 percent and the frontier one at 30.0. Both are true, about different classes of task. If the classes are 9x apart and misclassification cannot be repaired in flight, the 10x price spread is not the product — the classifier is. Two things worth carrying: the authors flag a hard-task exception where reduced-context escalation does win, on roughly 24 samples per cell, and call it exploratory; and the finding covers two model families. But the direction nobody is marketing looks close to free. Starting strong and finishing cheap raises the Claude pair from 54.6 to 65.6 percent while cost moves only $0.41 to $0.51, retaining 80 percent of the small model’s cost advantage. The cascade everyone built runs in the wrong direction.