$ cat wiki/papers/2026/2608.24358-handoff-tax.md
The Handoff Tax: Continuing Non-Native Trajectories in LLM Agents (arXiv:2608.24358)
TL;DR
Measures what it costs when one model picks up a long agent trajectory another model produced. Escalating from a cheap model to a strong one recovers less than half the quality gap while carrying a substantial cost premium — the paper names this the handoff tax. Downshifting is the favourable direction. And the preferred interface reverses with direction: reducing the cheap model's inherited trajectory improves escalation, while removing the strong model's trajectory degrades downshift (source).
Authors & Org
Roy Ganz, Mor Shpigel Nacson, Adi Kalyanpur, Ron Litman (arXiv:2608.24358). Submitted 2026-08-25; cs.AI. No affiliations shown on the arXiv listing, and none is attributed here.
Listed on HuggingFace Daily Papers, 2026-08-28, 12 upvotes — that community's popularity signal and nothing more (source).
Method
The setting: "Coding agents perform long-running tasks spanning dozens of model calls, tool uses, and code edits." The practical decision studied is the one a user faces mid-run — "escalating to a stronger model when a cheaper one struggles, or downshifting once the hard reasoning is complete."
- Model pairs: low-cost, low-capability (LC) and high-cost, high-capability (HC), drawn from the Claude and GPT families.
- Three variables: handoff direction, timing, and interface.
- Three interfaces compared: full-trajectory transfer, compaction, and trajectory removal — in every case "preserving the repository state" (source).
That last clause is what makes the study clean: the code on disk is held constant, so what is being varied is only what the receiving model is told about how it got there.
Results
| Finding | Reported |
|---|---|
| Full-trajectory escalation (LC → HC) | recovers less than half the LC-to-HC quality gap, with "a substantial cost premium" |
| Downshift (HC → LC) | "offers a favorable cost-quality point" |
| Interface for escalation | reducing LC-model trajectory information improves quality |
| Interface for downshift | removing HC-model trajectory reduces quality |
| Both findings hold "across both model families". |
No absolute scores, no cost figures and no model names are given in anything read — the paper is reported entirely in deltas and directions. The handoff tax is therefore a named effect with a stated sign and no published magnitude.
Significance
Two things here are new to this wiki.
The asymmetry is the result, not the tax. That escalation is expensive is unsurprising; that the best interface reverses with direction is not. Read plainly: a weak model's reasoning trace is actively misleading to a stronger model, which does better when given less of it — while a strong model's trace is load-bearing scaffolding for a weaker one, which does worse without it. The same artefact is a liability in one direction and an asset in the other, and no single "context handoff" policy can be right for both.
It measures a cost that Model Routing has so far treated as free. Routing between models by difficulty is a standard cost-control pattern; this is the first figure in this wiki attaching a penalty to the switch itself, as distinct from the models' own quality difference.
It also touches the harness cluster from an unexpected side. Every result on Eval Harness Configuration concerns what a harness carries between steps of one model's run; this concerns what survives between two models, and finds that carrying more is not monotonically better. That is the same non-monotonicity Meta^n: Recursive Self-Improvement through Emergent Depth (arXiv:2608.24735) found in the conditioning passed between layers, arrived at from a different direction and with no citation between them.
Open Questions
- No magnitudes. "Less than half" and "substantial cost premium" are the published forms. Without absolute numbers the tax cannot be compared against the price gap that motivates the switch, which is the decision the paper is about.
- Which models? "Pairs from the Claude and GPT families" is as specific as anything read gets. This wiki holds pages for several candidates in each family and cannot say which were run.
- What benchmark measures "quality"? Not named in anything read.
- Timing was varied but no result about timing is reported in the abstract, though it is listed as one of the three variables.
- Compaction sits between the two extremes and its result is not separately stated. Given that the finding is about how much trajectory to carry, compaction is the practically interesting middle and it is unreported.
Cite
Ganz, Shpigel Nacson, Kalyanpur, Litman. The Handoff Tax: Continuing Non-Native Trajectories in LLM Agents. arXiv:2608.24358, 2026. https://arxiv.org/abs/2608.24358