The Handoff Tax: Continuing Non-Native Trajectories in LLM Agents

Roy Ganz; Mor Shpigel Nacson; Adi Kalyanpur; Ron Litman

cs.AIAug 25, 2026

The Handoff Tax: Continuing Non-Native Trajectories in LLM Agents

Authors: Roy Ganz, Mor Shpigel Nacson, Adi Kalyanpur, Ron Litman

Organizations: AWS, Agentic AI

Abstract

Coding agents perform long-running tasks spanning dozens of model calls, tool uses, and code edits. As these runs unfold, users face a practical cost-quality trade-off: escalating to a stronger model when a cheaper one struggles, or downshifting once the hard reasoning is complete. Each switch requires the receiver to continue a non-native trajectory produced by another model. We study how this handoff affects quality and cost, and how varying the trajectory information inherited by the receiver changes the outcome. Using pairs of low-cost, low-capability (LC) and high-cost, high-capability (HC) models from the Claude and GPT families, we vary handoff direction, timing, and interface, comparing full-trajectory transfer, compaction, and trajectory removal while preserving the repository state. Across both model families, full-trajectory escalation recovers less than half of the LC-to-HC quality gap while incurring a substantial cost premium. We term this cost-quality penalty the handoff tax. By contrast, downshift offers a favorable cost-quality point. Interestingly, the preferred interface also reverses with direction: reducing LC-model trajectory information improves escalation quality, whereas removing the HC-model trajectory reduces downshift quality.

Figures & tables

Figure 1: The Handoff Tax. (a) Users escalate to a high cost model (HC) when the cheaper model (LC) struggles, or downshift to a cheaper one once hard reasoning appears complete. (b) At handoff, the receiver inherits another model’s trajectory. We ask how this affects cost and quality, and how varying the transferred trajectory information shapes the effect.

Figure 2: Handoff interfaces. All strategies preserve the prefix model’s edited working tree at the switch but vary the trajectory information transferred to the suffix: the full trajectory ( Raw ), a summary written by the prefix or suffix model ( Compact pre / Compact suf ), or none ( Traj-drop ).

Claude (Haiku 4.5 / Opus 4.7)						GPT-5.6 (Luna / Sol)
Strategy	Pass	Cost	Steps	QRec	CSRet	Strategy	Pass	Cost	Steps	QRec	CSRet
(a) Escalation (LC $\rightarrow$ HC)
LC-only	60.7	0.40	80	$0$	$100$	LC-only	58.7	0.06	14	$0$	$100$
HC-only	79.2	0.72	31	$100$	$0$	HC-only	83.7	0.47	17	$100$	$0$
Abort + HC fresh	79.2 †	0.90	78	$100^{\dagger}$	$-58$	Abort + HC fresh	83.7 †	0.51	27	$100^{\dagger}$	$-11$
LC-full + HC-full	79.2 †	1.12	111	$100^{\dagger}$	$-130$	LC-full + HC-full	83.7 †	0.53	31	$100^{\dagger}$	$-15$

Table 1: Aggregate coding-agent handoffs across model pairs. Raw escalation recovers less than half of HC’s quality advantage, whereas downshift offers a favorable intermediate cost–quality point. Reducing inherited LC context improves escalation, while preserving HC context improves downshift. Values are unweighted means over seven switch points on matched switched subsets; QRec and CSRet follow Sec. 3.4 .

Figure 3: Handoff effects vary with task difficulty. For Claude, all escalation interfaces are unfavorable on easy and medium tasks, but on hard tasks reduced-context interfaces become cheaper than HC-only while recovering 65–74% of HC’s quality advantage; Raw does not make this transition. In downshift, dropping the HC trajectory yields the worst cost–quality trade-off among tested interfaces in every difficulty bucket.

Figure 4: The handoff tax arises through different computational mechanisms. In Claude escalation, Raw makes each post-handoff HC step

2.2\times

costlier than Compact pre ; in downshift, Traj-drop makes LC take

1.6\times

more steps. Full LC context inflates HC calls, whereas missing HC context forces LC rework.

Figure 5: Task-relevant information evolves differently across settings. SWE-bench provides the specification upfront, followed by rapid repository scoping; LiC reveals user requirements over turns; BrowseComp accumulates evidence through web search. The schematic illustrates what information is available at different stages of execution.

Tasks	Policy	Score	Cost	QRec	CSRet
[-0.1em]	LC-only	62.1	0.015	$0$	$100$
	HC-only	76.9	0.080	$100$	$0$
	Escalation	74.2	0.056	$86$	$36$
	Downshift	67.0	0.046	$31$	$53$

Table 2: Late-arriving requirements favor an HC suffix. Raw-handoff results on LiC, averaged across five task families and early, middle, and late structural switch positions for Claude.

Strategy	QRec	CSRet	Steps
			LC	HC
Full LC + full HC	100.0	$-40.1$	62.6	33.4
Abort + fresh HC	100.0	$-8.5$	23.0	33.4
Escalation	95.8	$-30.0$	23.0	30.4
Downshift	56.7	76.8	22.8	17.0

Table 3: Under progressive search, escalation recovers quality but not savings. GPT Raw-handoff results on BrowseComp, averaged across switch points.

Appendix figures & tables8 assets

Supplementary material from the paper’s appendix.

Appendix

Figure 6: GPT exhibits the same handoff-cost accounting pattern as Claude. Relative to Compact pre , each post-handoff HC step under Raw costs

1.6\times

as much, while Traj-drop downshift takes

2.0\times

as many post-handoff LC steps.

Claude (Haiku 4.5 / Opus 4.7)						GPT-5.6 (Luna / Sol)
Strategy	Pass	Cost	Steps	QRec	CSRet	Strategy	Pass	Cost	Steps	QRec	CSRet
All
LC-only	60.7	0.40	80	$0$	$100$	LC-only	58.7	0.06	14	$0$	$100$
HC-only	79.2	0.72	31	$100$	$0$	HC-only	83.7	0.47	17	$100$	$0$
Abort + HC fresh	79.2 †	0.90	78	$100^{\dagger}$	$-58$	Abort + HC fresh	83.7 †	0.51	27	$100^{\dagger}$	$-11$
LC-full + HC-full	79.2 †	1.12	111	$100^{\dagger}$	$-130$	LC-full + HC-full	83.7 †	0.53	31	$100^{\dagger}$	$-14$

Appendix

Table 5: Coding-agent escalation by switch-timing window. Columns compare model families. All averages p5–p50, Early averages {p5,p10,p15}, and Late averages {p25,p35,p45,p50}. Values are unweighted means over switch points on matched switched subsets. Pass is in percent, Cost in dollars, and Steps is the mean trajectory length. QRec is recovered LC–HC quality gap; CSRet is retained cost savings, with negative values costlier than HC-only. Gray rows are single-model baselines and restart references. Bold marks the best available handoff quality and best non-baseline cost/efficiency within each family and timing window.

Claude (Haiku 4.5 / Opus 4.7)						GPT-5.6 (Luna / Sol)
Strategy	Pass	Cost	Steps	QRec	CSRet	Strategy	Pass	Cost	Steps	QRec	CSRet
All
LC-only	54.6	0.41	79	$0$	$100$	LC-only	63.6	0.05	13	$0$	$100$
HC-only	75.8	0.85	35	$100$	$0$	HC-only	85.8	0.47	18	$100$	$0$
Raw handoff	65.6	0.51	58	$50$	$\mathbf{80}$	Raw handoff	81.0	0.42	16	$\mathbf{79}$	$\mathbf{14}$
Compact ${}_{\text{pre}}$	66.8	0.52	54	$\mathbf{56}$	$78$	Compact ${}_{\text{pre}}$	79.8	0.43	18	$72$	$10$

Appendix

Table 6: Coding-agent downshift by switch-timing window. Columns compare model families. All averages p5–p50, Early averages {p5,p10,p15}, and Late averages {p25,p35,p45,p50}. Values are unweighted means over switch points on matched switched subsets. Pass is in percent, Cost in dollars, and Steps is the mean trajectory length. QRec is recovered LC–HC quality gap; CSRet is retained cost savings. Gray rows are single-model baselines. Bold marks the best available handoff quality and best non-baseline cost/efficiency within each family and timing window.

Strategy	Pass	Cost	Steps	QRec	CSRet
	(%)	($)	(#)	(%)	(%)
LC-only	60.6	0.41	80	$0$	$100$
HC-only	78.8	0.72	31	$100$	$0$
Raw handoff	69.3	1.62	74	$48$	$-285$
Raw + disclosure	70.5	1.67	75	$55$	$-300$
Compact ${}_{\text{pre}}$	71.5	0.75	71	$60$	$\mathbf{-10}$

Appendix

Table 7: Escalation (LC

\rightarrow

HC), averaged over all switch points. Intersection-only results average p5–p50 unweighted. QRec is quality-gap recovery and CSRet is cost-savings retention; both are 100% at the ideal corner. LC =Haiku 4.5 and HC =Opus 4.7. Compact

{}_{\text{pre}}

/Compact

{}_{\text{suf}}

use the handoff-from/handoff-to model to summarize. Raw + disclosure retains the full trajectory and appends one model-change notice at the first HC call. Best Pass/QRec among handoff strategies and best Cost/Steps/CSRet among all non-baseline methods are in bold .

Direction	Comparison		Claude		GPT
Direction	First	Second	$\Delta$ Pass	95% CI	$\Delta$ Pass	95% CI
Escalation	Traj-drop	Raw	$+3.1$	$[+1.3,+5.0]$	$+12.2$	$[+9.2,+15.4]$
	Traj-drop	Compact ${}_{\text{pre}}$	$+0.6$	$[-1.1,+2.3]$	$+10.9$	$[+7.5,+14.5]$
	Traj-drop	Compact ${}_{\text{suf}}$	$+2.8$	$[+1.0,+4.6]$	$+10.9$	$[+7.7,+14.2]$
	Compact ${}_{\text{pre}}$	Raw	$+2.5$	$[+0.9,+4.2]$	$+1.3$	$[-1.7,+4.3]$
	Compact ${}_{\text{suf}}$	Raw	$+0.4$	$[-1.2,+1.9]$	$+1.3$	$[-1.4,+4.0]$

Appendix

Table 8: Pairwise interface pass-rate differences by model family.

\Delta

Pass is the first strategy minus the second, in percentage points. Within each direction, rows present the central Traj-drop–Raw reversal, broader Traj-drop–compaction comparisons, and secondary comparisons among context-preserving interfaces, in that order. Intervals are pointwise 95% task-clustered bootstrap confidence intervals from

10{,}000

resamples.

Task	Policy	Early				Middle				Late
		Score	Cost	QRec	CSRet	Score	Cost	QRec	CSRet	Score	Cost	QRec	CSRet
			($)	(%)	(%)		($)	(%)	(%)		($)	(%)	(%)
Code	LC-only	65.0	0.033	$0$	$100$	65.0	0.033	$0$	$100$	65.0	0.033	$0$	$100$
	HC-only	92.0	0.163	$100$	$0$	92.0	0.163	$100$	$0$	92.0	0.163	$100$	$0$
	Escalation	88.0	0.145	$85$	$14$	89.0	0.133	$89$	$23$	80.0	0.065	$56$	$75$
	Downshift	75.0	0.047	$37$	$89$	85.0	0.082	$74$	$62$	85.0	0.156	$74$	$5$

Appendix

Table 9: Raw handoff by switch position in sharded, underspecified conversations, Claude pair. Early, middle, and late are structural switch positions. Later escalation retains more savings, while quality is mostly stable unless prior turns accumulate deliverable state. Later downshift generally improves quality while retaining less savings. Score is pass rate except for Data-to-text, which reports BLEU

\times 100

Task	Policy	Early				Middle				Late
		Score	Cost	QRec	CSRet	Score	Cost	QRec	CSRet	Score	Cost	QRec	CSRet
			($)	(%)	(%)		($)	(%)	(%)		($)	(%)	(%)
Code	LC-only	85.0	0.017	$0$	$100$	85.0	0.017	$0$	$100$	85.0	0.017	$0$	$100$
	HC-only	95.0	0.057	$100$	$0$	95.0	0.057	$100$	$0$	95.0	0.057	$100$	$0$
	Escalation	96.0	0.052	$110$	$13$	94.0	0.037	$90$	$49$	94.0	0.024	$90$	$82$
	Downshift	90.0	0.024	$50$	$82$	94.0	0.037	$90$	$50$	96.0	0.052	$110$	$13$

Appendix

Table 10: Raw handoff by switch position in sharded, underspecified conversations, GPT pair. Early, middle, and late are structural switch positions. Later escalation retains more savings, while quality is mostly stable unless prior turns accumulate deliverable state. Later downshift generally improves quality while retaining less savings. Score is pass rate except for Data-to-text, which reports BLEU

\times 100

Prefix model	Difficulty	$N$	Switch step by prefix-run percentile
Prefix model	Difficulty	$N$	p5	p10	p15	p25	p35	p45	p50
Haiku 4.5	Easy	194	25	30	34	37	42	46	50
	Medium	261	35	41	45	51	58	63	66
	Hard	45	48	54	58	63	77	86	94
Opus 4.7	Easy	194	5	7	8	10	12	14	14
	Medium	261	8	12	13	16	18	21	23

Appendix

Table 11: Difficulty-calibrated switch steps by prefix model. Steps are agent turns, corresponding to model API calls. Each threshold is the specified percentile of the prefix model’s single-model termination-step distribution within the given difficulty bucket.

LLM Agents

Large language model agents use language models to plan, use tools, and act across multi-step tasks and environments.

See all 681 papers

TACIT-Switch: Cost-Aware Model Escalation for LLM Agents from Censored Supervision
Aug 28, 2026Ji'an Lei, Jian HuangLLM Agent Routing LLM Agent Reliability
School of Statistics Beijing Normal University Beijing 100875, China · Department of Applied Mathematics The Hong Kong Polytechnic University Hung Hom, Kowloon, Hong Kong SAR, China
Attention Tax, Handoff Tax: A Stylised Model of When Multi-Agent LLM Systems Help
Oct 5, 2026Akshit Anchan, Nayonika SenMulti-Agent LLM Systems Multi-Agent Reasoning
University of Amsterdam Amsterdam, The Netherlands · Northeastern University Boston, MA, United States of America
Whose turn is it? Context-guided model routing for stateful agents
Oct 13, 2026Anikesh Parashar, Avik Dutta, Vashu Chauhan +6LLM Routing LLM Agent Routing
Microsoft

Aug 28, 2026J/K move · Enter open · S save

TACIT-Switch: Cost-Aware Model Escalation for LLM Agents from Censored Supervision

Ji'an Lei, Jian Huang

School of Statistics Beijing Normal University Beijing 100875, China · Department of Applied Mathematics The Hong Kong Polytechnic University Hung Hom, Kowloon, Hong Kong SAR, China

LLM Agent Routing LLM Agent Reliability

Open paper AskPDF arXiv

Agents with smaller language-model backbones are less expensive but can drift into persistent failure modes, whereas those with larger backbones are generally more reliable but more costly. This reliability-cost trade-off motivates routing methods that decide when to invoke an agent with a larger backbone: before execution, after a fixed trajectory prefix, or locally at individual steps. Our method, TACIT-SWITCH, learns permanent handoff policies from accumulated trajectory evidence and Teacher-Annotated Censored Intervention Times (TACIT). It represents each annotation as an interval-censored observation on a cumulative-risk scale. The resulting mixture-cure threshold model estimates the probability that the paired Strong rollout succeeds and, conditional on success, the handoff threshold; no teacher is required at deployment. In a mechanism-based multi-step simulation, TACIT-SWITCH improves success by 7.4-11.1 percentage points over task-level, step-level, and fixed-prefix routing baselines at comparable cost. Within that controlled simulation, ablations show that task features and cumulative trajectory risk provide complementary information. With operating points selected on development data, TACIT-SWITCH achieves the highest held-out success among learned policies on both ALFWorld (48.5% with 4B Cheap; 45.5% with 9B Cheap) and DABench (73.1%).

Abstract

LLM Agents

Explore similar work

TACIT-Switch: Cost-Aware Model Escalation for LLM Agents from Censored Supervision

Attention Tax, Handoff Tax: A Stylised Model of When Multi-Agent LLM Systems Help

Whose turn is it? Context-guided model routing for stateful agents