Pack Light, CarryOn: Building Object-Oriented Context for Long-Horizon Agents

Yuchen Lu; Julian McAuley; Lei Zhang

cs.AIOct 13, 2026

Pack Light, CarryOn: Building Object-Oriented Context for Long-Horizon Agents

Authors: Yuchen Lu, Julian McAuley, Lei Zhang

Organizations: University of California San Diego

Abstract

Long-horizon agents accumulate progress in persistent task state, yet repeatedly process interaction histories describing how that progress was achieved. This raises a natural question: can long-horizon context instead be organized around reusable progress rather than the history that produced it? In this paper, we introduce CarryOn, an object-oriented context runtime that organizes continuation around reusable task objects. Rather than treating compressed histories or reconstructed accounts of progress as the primary basis for continuation, CarryOn builds a context view from continuation carriers, complementary execution context, and a live tail of recent interactions. A core challenge is keeping this reusable basis valid after its surrounding history is retired: how can CarryOn update the basis without reconstructing the full representation at every fold? CarryOn maintains carrier records across folds, revising affected records against bound evidence and applicable environment checks while carrying forward unaffected records and jointly updating execution context. Experiments on OSWorld 2.0 and BrowseComp-Plus show improved gains of 5.8% and 7.0% while reducing total token usage by 51.9% and 22.9%, respectively. Project page: https://ethannju.github.io/carryon

Figures & tables

Figure 1: Context folding with reusable task objects in long-horizon OSWorld2.0 ( Yuan et al., 2026 ) . Top: In a slide-editing task, the working presentation and exported images carry completed editing progress, while remaining constraints and save verification stay explicit, reducing active context from 42.4K to 14.4K tokens (66.0%). Bottom: In a lab-report task, the working document carries completed edits, while unresolved calculations remain explicit, reducing active context from 58.4K to 23.5K tokens (59.7%).

Figure 2: Overview of the CarryOn framework. CarryOn maintains an object-oriented context view comprising continuation carriers, execution context, and a live tail. When invoked, it tracks reusable progress, binds supporting evidence, checks relevant environment conditions, and jointly compiles carrier and execution-context updates. A successful fold commits validated updates and retires covered live-tail interactions from active context, while interaction log remains append-only.

Model / thinking effort	Binary (%) $\uparrow$	Partial (%) $\uparrow$	Cost/task ($)
GPT-5.5 (xhigh) ( Yuan et al., 2026 )	13.0	49.5	25.5
Claude Opus 4.8 (max) ( Yuan et al., 2026 )	18.5	49.3	76.1
Claude Sonnet 4.6 (max) ( Yuan et al., 2026 )	8.3	41.5	22.3
Claude Sonnet 4.6 (medium) ( Yuan et al., 2026 )	9.3	33.9	14.4
MiniMax M3 (enabled) ( Yuan et al., 2026 )	4.6	22.3	2.4
Kimi 2.6 (enabled) ( Yuan et al., 2026 )	4.6	22.1	6.6

Table 1: Task performance on OSWorld 2.0. ∗ means we report the reproduced results using the official benchmark and API. CarryOn improves both binary (+3.7%) and partial (+5.8%) task scores over the same Claude Opus 4.7 (max) host agent while substantially reducing cost per task (-35.9%).

	OSWorld 2.0		BrowseComp-Plus
Metric	Baseline	CarryOn	Baseline	CarryOn
Agent tokens (M)	40.10	18.61 $\downarrow$ 53.6%	1.56	1.17 $\downarrow$ 24.6%
Runtime tokens (M)	–	0.69	–	0.03
Total tokens (M)	40.10	19.30 $\downarrow$ 51.9%	1.56	1.20 $\downarrow$ 22.9%
Agent cost ($)	36.83	20.51 $\downarrow$ 44.3%	1.89	1.62 $\downarrow$ 14.4%
Runtime cost ($)	–	3.11	–	0.14

Table 2: Evaluation of resource usage on OSWorld 2.0 and BrowseComp-Plus. Values are averaged per task. CarryOn reduces total token usage (by 51.9% and 22.9% respectively) and cost (by 35.9% and 7.1% respectively) on both benchmarks.

Model	Retriever	Accuracy (%) $\uparrow$
OpenAI GPT-5 ( Chen et al., 2025 )	BM25	52.00
OpenAI o3 ( Chen et al., 2025 )	BM25	50.00
OpenAI GPT-OSS-120B ( OpenAI et al., 2025 )	BM25	41.00
Claude Opus 4.7	BM25	49.00
Claude Opus 4.7 + CarryOn	BM25	56.00

Table 3: Task performance on the 100-question subset of BrowseComp-Plus. CarryOn improves the accuracy over the same Claude Opus 4.7 by 7%.

Figure 3: Context and cost dynamics on OSWorld 2.0. (a) Mean input context size and (b) mean cumulative model cost over agent turns. CarryOn keeps input context near 65K tokens while the baseline continues to grow, and its cost advantage widens with trajectory length, reaching a 34.1% reduction by turn 500. Gray bars indicate the fraction of baseline trajectories remaining active; costs include context-management overhead.

Method	Text Length (k chars)	Binary (%) $\uparrow$	Partial (%) $\uparrow$
Baseline	–	22.50	50.32
Sliding Window	12.08	15.00	37.41
Generic Summary (GS)	17.60	15.00	44.94
CarryOn-Guided Summary (CGS)	23.94	20.00	50.93
CarryOn (ours)	21.54	22.50	55.93

Table 4: Comparison of different methods on a shared 40-task subset of OSWorld 2.0. Post-fold text length measures the context retained after each compression or fold. CarryOn achieves the highest partial score (55.93%) without retaining the most text.

Method	Agent Calls	Cost ( $)$ \downarrow$	Binary (%) $\uparrow$	Partial (%) $\uparrow$
w/o Execution Context	274.33	18.33	17.50	44.00
w/o Continuation Carriers	349.45	25.14	25.00	44.62
CarryOn (ours)	282.25	20.55	22.50	55.93

Table 5: Context-view ablation on the shared 40-task subset of OSWorld 2.0. Each ablation hides one component from the host agent while preserving CarryOn’s internal runtime state. Removing either component substantially lowers partial score, while removing continuation carriers also increases agent calls and cost.

Figure 4: Carrier persistence and continuation after folding. (a) Across committed folds on OSWorld 2.0, 80% of carrier observations are retained or updated rather than newly added. (b) Representative trajectories show the agent continuing directly from reusable task state after earlier interactions have been folded, without repeating the work that produced it.

Appendix figures & tables9 assets

Supplementary material from the paper’s appendix.

Appendix

Check	Observation
File existence	Whether the file is present at its bound path.
Change from initial state	Whether the file-content hash differs from its registered initial snapshot.
File size	Whether the file is empty or nonempty.
Document structure	Whether a docx , pptx , or xlsx file has a valid document container.
Document differences	Changes in document text units, added or removed parts, and placeholder-like content.
Instruction text matching	Presence or absence of a text span copied from the public task instruction.

Appendix

Table 6: Read-only environment checks in the OSWorld 2.0 integration.

Setting	OSWorld 2.0	BrowseComp-Plus
Agent execution budget	500 steps	60 turns
Agent output limit	16,000	32,000
Tracker output limit	2,048	8,192
Compiler output limit	8,192	8,192
Compilation requests per scan	4	4

Appendix

Table 7: Agent and CarryOn settings. Output limits are measured in tokens per model call.

Token category	Price ($/M tokens)
Uncached input	5.00
Cache read	0.50
Cache write (5 minutes)	6.25
Cache write (1 hour)	10.00
Output	25.00

Appendix

Table 8: Nominal token prices used for resource accounting.

Step	Recorded state or agent statement	Action
203	The agent reports that all four external authors have been entered.	Continues through the form.
205	Publication metadata is entered alongside the author list.	Types the NeurIPS conference dates.
F4 after 240	Carrier: entered authors and publication metadata. EC: classify authors by affiliation at publication time.	Installs the updated context view.
247	Reviews the existing author classifications against the guidelines.	Presses Escape after the review.
249	The publishing status, peer-review status, and attached PDF remain in the form.	Returns to the form without re-entering the authors.
250	The existing form is ready for submission.	Clicks Save & Submit; the page confirms successful submission.

Appendix

Table 9: Continuing from a populated form (Task 009). Selected steps from a main-experiment CarryOn trajectory. Green highlights continuation from the existing form. Entries are condensed from the recorded trajectory; the submission confirmation is also visible in the saved screenshot.

Step	Recorded state or agent statement	Action
CarryOn
30	Identifies all 15 catalogue items and their prices, then develops a prize-distribution plan.	Closes the catalogue viewer.
F2 after 57	Carrier: item-price mapping and chosen tier assignments. EC: enter the third-prize total, then complete the participation tier.	Both records remain visible to the agent.
58	Continues entering the planned third-prize total.	Types 125 in the total field.
60	Moves on to the planned participation-prize items.	Selects Chocolate Box.
w/o Carrier

Appendix

Table 10: Continuing a plan versus reacquiring its inputs (Task 013). Selected steps from the full CarryOn and carrier-masking runs. Green marks continuation of the retained prize plan; red marks renewed uncertainty and repeated acquisition of the price mapping. Quoted spans are exact excerpts from recorded agent reasoning summaries; other descriptions are condensed.

System prompt

You are the Carrier Tracker and Readiness Judge of a Context Runtime for a

computer-use agent. You observe the agent’s interaction stream step by step and

hold TWO equally weighted duties:

DUTY A — LEDGER: maintain a sparse identity ledger of the continuation carriers

the agent is producing (the candidates).

Appendix

Table 11: Tracker prompt. Selected instructions from the routine Tracker system prompt, adapted for presentation. Ellipses mark omitted passages.

Environment objects

You are the TYPED ENVIRONMENT CARRIER AND RESIDUAL-EC COMPILER of a Context

Runtime. A machine Grounder has authorized one ordered vector of external_object

and/or situated_state Candidates. Author exactly those typed carriers plus a

compact residual execution context that, together, can replace the public Live

Tail being folded.

Appendix

Table 12: Compiler prompts. Selected instructions from the three alternative Compiler system prompts, adapted for presentation. Route headings are added; ellipses mark omitted passages.

System prompt

Read the whole conversation so far carefully — every earlier message in order,

and the previous summary if one is present — and summarize it so that the

assistant can continue working on the task from this summary together with the

most recent messages. Be thorough and go into detail: work through the history

from beginning to end and cover what actually happened, rather than writing a

short or high-level gist. Output only the summary text, with no tool calls. Aim

Appendix

Table 13: Generic Summary (GS) prompt. Complete system prompt used for Generic Summary (GS) in the context-management comparison.

System prompt

You maintain a detailed, self-contained continuation summary for an assistant

working on a long-running task. The assistant will receive your updated summary

together with a retained tail of recent native messages and available

observations. Much of the earlier conversation may no longer be directly

available after this update.

Appendix

Table 14: CarryOn-Guided Summary (CGS) prompt. Selected instructions from the CarryOn-Guided Summary (CGS) system prompt. Ellipses mark omissions.

LLM Context Management

Large language model context management studies how to select, compress, retrieve, and organize information within bounded model inputs.

See all 248 papers

StateComp: Learning When to Compress History in Long Horizon Agents
Sep 23, 2026Mingxuan Wang, Hongyue Chen, Yinglong Guo +6Context Compression Long-Horizon LLM Agents
TierFlow Team · Gaoling School of Artificial Intelligence, Renmin University of China · Tsinghua University
LongSeeker: Elastic Context Orchestration for Long-Horizon Search Agents
May 6, 2026Yijun Lu, Rui Ye, Yuwen Du +3Web Search Agents Tool-Using Agents
Shanghai Jiao Tong University
ContextRender: From Execution Dependencies to Agent Context
Sep 29, 2026Savini Kashmira, Jayanaka L. Dantanarayana, Lingjia Tang +1Graph-Based Agent Memory Long-Horizon LLM Agents
University of Michigan, Ann Arbor, USA

Sep 23, 2026J/K move · Enter open · S save

StateComp: Learning When to Compress History in Long Horizon Agents

Mingxuan Wang, Hongyue Chen, Yinglong Guo, Fei Luo+5

TierFlow Team · Gaoling School of Artificial Intelligence, Renmin University of China · Tsinghua University

Context Compression Long-Horizon LLM Agents Agent Memory

Open paper AskPDF arXiv

Long-horizon agents continuously accumulate interaction history during task execution, yet the importance of past interactions changes as the agent state evolves. Existing context management methods largely compress history based on fixed windows, periodic schedules, or current relevance, overlooking a more fundamental question: when has a past interaction become safe to replace? Premature compression may remove information still needed for future actions, while overly conservative retention leads to substantial context overhead. To address this, we propose State Conditioned Compression (StateComp), a framework that determines when historical interactions can be safely compressed according to the current agent state. StateComp constructs KEEP and READY supervision through a two-stage annotation procedure and trains an imbalance-aware router on hidden representations from a frozen language model. A bounded state representation further reduces the cost of evaluating long histories, while adjacent READY interactions are grouped into continuous spans and replaced with compact summaries during execution. Experiments on WorkBuddyBench show that StateComp reduces total agent and summarization tokens by 52.27% while maintaining task performance, and achieves a 12.67-fold speedup in representation extraction.

Abstract

LLM Context Management

Explore similar work

StateComp: Learning When to Compress History in Long Horizon Agents

LongSeeker: Elastic Context Orchestration for Long-Horizon Search Agents

ContextRender: From Execution Dependencies to Agent Context