PaperlayerView exact revision on arXiv

Source-linked AI summary

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

Yuchen Lu, Julian McAuley, Lei Zhang

arXiv:2610.18784v2cs.AIcs.CL

TL;DR

Long-horizon agents repeatedly process interaction histories describing progress already captured in persistent task state. CarryOn organizes continuation around reusable task objects, complementary execution context, and recent interactions; it improves performance on OSWorld 2.0 and BrowseComp-Plus while reducing token use.

  • Problem

    Long-horizon agents repeatedly process interaction histories describing progress already embodied in persistent task state, and history-based approaches can make continuation depend on accounts that drift from the state agents inspect and modify.

    Source: §1 · p. 1

  • Method

    CarryOn combines reusable continuation carriers, execution context for active goals and unresolved issues, and a live tail, updating carrier records against evidence while preserving unaffected records.

    Source: §1 · pp. 1–2

  • Results

    CarryOn raises Claude Opus 4.7 from 43.9% to 49.7% (+5.8%) on OSWorld 2.0 and from 49% to 56% (+7%) on BrowseComp-Plus, reducing total token usage by 51.9% and 22.9%, respectively.

    Source: §1 · p. 2, §4.2 · p. 7

  • Takeaways & Limitations

    Across computer-use and information-seeking tasks, CarryOn reduces context-processing overhead while maintaining or improving observed task performance.

    Source: §4.2 · p. 7

Abstract

from arXiv · show

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

1 INTRODUCTION

CarryOn organizes long-horizon context around reusable task objects rather than accounts of how progress was made. It combines those objects with complementary execution context and recent interactions, improving benchmark performance while reducing token use.

1 INTRODUCTION

  • Long-horizon agents repeatedly process histories of progress already embodied in persistent task state, while history-based approaches can leave continuation dependent on accounts that drift from that state.Prior approaches summarize, retrieve, or prune interactions, or translate them into progress records and structured task states.Source: §1 · p. 1
  • CarryOn makes reusable task objects the basis for continuation, pairing them with execution context for active goals and unresolved issues and a live tail of recent interactions.Carriers bind reusable content or object references to their conditions of use.Source: §1 · p. 1
  • CarryOn revises affected carrier records against supporting evidence and applicable environment checks, carries forward unaffected records, and jointly updates execution context as history is folded.Execution context preserves goals, constraints, and unresolved issues not captured by carriers.Source: §1 · p. 2
  • CarryOn lifts Claude Opus 4.7 from 43.9% to 49.7% (+5.8%) on OSWorld 2.0 and from 49% to 56% (+7%) on BrowseComp-Plus, reducing total token usage by 51.9% and 22.9%, respectively.Total cost falls by 35.9% and 7.1% on the two benchmarks, respectively.Source: §1 · p. 2

2 RELATED WORK

Prior long-horizon context methods either manage interaction histories or externalize task progress. CarryOn instead makes reusable task objects the primary unit of continuation.

2 RELATED WORK

  • Context-management methods compress, prune, summarize, or selectively retrieve prior interactions, while external-history methods bound active context but retain access to earlier interactions.Other approaches validate compressed representations against subsequent execution to reduce information loss.Source: §2 · p. 3
  • State-externalization methods maintain internal state, structured task representations, persistent workspaces, or reconstructed working contexts.These approaches organize continuation around compressed interaction history or externally maintained state.Source: §2 · p. 3
  • CarryOn differs by making reusable task objects, rather than compressed histories or externally maintained state, the primary unit of continuation.This contrast is stated across the surveyed approaches.Source: §2 · p. 3

3 METHOD

CarryOn organizes continuation around reusable task objects, complementary execution context, and recent interactions. It folds interaction history into this maintained view by updating selected progress records and preserving information needed to continue.

3.1 FRAMEWORK

  • The framework tracks candidate reusable progress, binds it to evidence and applicable environment checks, jointly compiles carrier and execution-context updates, then folds covered interactions from active context.Subsequent interactions form a new live tail.Source: §3.1 · p. 3

3.2 CONTEXT VIEW

  • The context view comprises continuation carriers for reusable task objects, execution context for complementary continuation information, and a live tail of interactions since the latest fold.A task may use external objects, information products, and situated state as carriers; these types can coexist and be revised.Source: §3.2 · pp. 3–4
  • Execution context retains current focus, unresolved constraints and obligations, and pending work not materialized into a carrier; the live tail keeps post-fold interactions directly available.Each fold maintains carrier records and execution context while the live tail contains interactions after the last covered step.Source: §3.2 · p. 4

3.3 CONSTRUCTING AND UPDATING THE CONTEXT VIEW

  • CarryOn tracks potential new or updated carriers across invocations and compiles candidates when ready, using authorized sources, prior carrier records, and available environment-check results as evidence.Candidates not yet ready remain tracked for later folds.Source: §3.3 · pp. 4–5
  • The compiler edits selected carriers against authorized evidence while revising execution context against the entire live tail, preserving unfinished requirements and carrying unaffected records forward.Validation checks versions, evidence consistency, and required constraints before committing; these checks do not verify task correctness.Source: §3.3 · p. 5
  • On OSWorld 2.0, CarryOn improves binary task score by 3.7% and partial task score by 5.8% over the same Claude Opus 4.7 (max) host agent, while reducing per-task cost by 35.9%.Across OSWorld 2.0 and BrowseComp-Plus, CarryOn reduces average total tokens per task by 51.9% and 22.9%, respectively, and cost by 35.9% and 7.1%, respectively.Source: Table 1 · p. 6, Table 2 · p. 6

4 EXPERIMENT

Experiments across computer-use and browsing tasks assess CarryOn’s performance and context costs, then examine context dynamics, alternative context-management strategies, and the roles of its context-view components.

4.1 EXPERIMENTAL SETUP

  • Evaluation uses full OSWorld 2.0 and 100 sampled BrowseComp-Plus questions, Claude Opus 4.7, and efficiency measures that include agent and context-management calls.OSWorld tasks have a 500-step budget; BrowseComp-Plus uses BM25 document retrieval.Source: §4.1 · p. 6

4.3 CONTEXT AND COST DYNAMICS

  • On OSWorld 2.0, CarryOn holds mean input context near 65K tokens while the baseline exceeds 200K, and cumulative cost falls 34.1% by turn 500 after crossing before turn 200.The cost reduction includes context-management overhead.Source: Figure 3 · p. 7, §4.3 · pp. 7–8

4.2 MAIN RESULT

  • CarryOn raises OSWorld 2.0 partial score from 43.9% to 49.7% and BrowseComp-Plus accuracy from 49.0% to 56.0%, while cutting total tokens 51.9% and 22.9%, respectively.On OSWorld 2.0, total cost per task falls 35.9%; on BrowseComp-Plus, total cost falls 7.1%.Source: §4.2 · p. 7

4.4 COMPARISON WITH CONTEXT MANAGEMENT STRATEGIES

  • Sliding Window and Generic Summary shorten contexts but reduce scores; CarryOn-Guided Summary reaches 50.93% partial score, while CarryOn reaches 55.93% with about 10% shorter post-fold text.CarryOn matches the baseline binary score of 22.50%.Source: §4.4 · p. 8

4.5 ANALYSIS

  • Removing execution context lowers binary score from 22.50% to 17.50% and partial score from 55.93% to 44.00%; removing carriers lowers partial score to 44.62% and raises calls 23.8%.The execution-context ablation changes agent-call count little, while removing carriers increases cost by 22.3%.Source: §4.5 · p. 8
  • Across committed folds, 80% of carrier records are retained or updated; representative trajectories continue from established task state without repeating the work that produced it.CarryOn’s explicit record maintenance contrasts with CGS, which generates a complete replacement summary at each fold; CarryOn scores 55.93% versus CGS’s 50.93% partial score.Source: §4.5 · pp. 8–9, Figure 4 · p. 9

5 CONCLUSION

CarryOn organizes long-horizon context around reusable task objects, complementary execution context, and recent interactions. On two benchmarks, it improves task scores while reducing token usage and cost.

5 CONCLUSION

  • CarryOn combines continuation carriers, execution context, and a live tail, jointly updating carriers and context as task state and remaining requirements evolve.Covered interaction history is progressively replaced by the updated context view.Source: §5 · p. 9
  • On OSWorld 2.0, CarryOn raises partial score from 43.9% to 49.7% while reducing total token usage by 51.9% and nominal model cost by 35.9%.The cost reduction includes runtime overhead.Source: §5 · p. 9
  • On BrowseComp-Plus, CarryOn increases accuracy from 49% to 56% with lower token usage and cost.Source: §5 · p. 9
  • The results support object-oriented context management in which task objects carry established progress and explicit context preserves what remains needed to continue.Source: §5 · p. 9

A METHOD AND IMPLEMENTATION DETAILS

The method maintains versioned reusable-progress records alongside execution context, then folds covered interactions into that representation through evidence-bound updates and validation. Its benchmark adapters install committed context and use different scan schedules and integration procedures.

A.1 CONTEXT REPRESENTATION AND UPDATE PROCEDURE

  • Committed carrier records and execution-context revisions define the folded context view, while later interactions remain in a live tail.Source: §A.1 · p. 13
  • Candidates can persist across Tracker invocations, accumulating supporting interaction references before compilation; registering or revising one does not trigger a fold.Each candidate records a provisional carrier type, description, identity fields, and observation and revision times.Source: §A.1 · p. 13
  • Carrier updates create linked successor versions, preserve unchanged records, and can add, modify, or retract information-product claims.Execution context is revised with carrier updates to retain unresolved requirements and remove resolved or carrier-represented items.Source: §A.1 · p. 13

A.2 GROUNDER CHECKS AND VERDICT RULES

  • Grounder checks bind requests to candidate identity, object reference, predicate, and expected value, then classify conditions as supported, contradicted, or unknown.Information products receive source-reference and excerpt checks; object-bound requests can receive scoped, read-only environment checks.Source: §A.2 · pp. 13–14, Table 6 · p. 14
  • A request proceeds only when all declared conditions are supported; unsupported or unknown checks leave that candidate uncompiled, while other admissible requests may proceed.These checks are consistency safeguards, not claims of task correctness.Source: §A.2 · p. 14

A.3 AGENT INTEGRATION AND CARRYON SETTINGS

  • Adapters install committed records and execution context before the next agent call; OSWorld scans after an estimated 6,000 tokens, while BrowseComp-Plus schedules scans by prompt-size growth.A failed context installation restores a snapshot when possible; BrowseComp-Plus scans back off after repeated scans without a fold.Source: §A.3 · pp. 14–15

B.1 BENCHMARK AND TASK SELECTION

The evaluation uses long-horizon computer-use tasks from OSWorld 2.0 and iterative-search questions from BrowseComp-Plus. The main evaluation covers the full OSWorld set, while the described comparison and BrowseComp-Plus experiments use subsets.

B.1 BENCHMARK AND TASK SELECTION

  • OSWorld 2.0 spans seven professional domains and 21 subcategories, with workflows combining desktop applications and web services and a median estimated human completion time of about 1.6 hours.The evaluation uses the v2026.08.08 release.Source: §B.1 · p. 15
  • The main OSWorld 2.0 evaluation runs the baseline and CarryOn on the full evaluation set; additional context-management comparisons and ablations use a shared 40-task subset.Source: §B.1 · p. 15
  • BrowseComp-Plus has 830 questions over 100,195 documents; the experiments evaluate the baseline and CarryOn on the same 100-question subset using BM25 retrieval.Its questions require iterative search and reasoning across multiple sources.Source: §B.1 · pp. 15–16

B.2 PREVIOUSLY REPORTED RESULTS

The paper uses published OSWorld 2.0 results and separately computed BrowseComp-Plus reference results for comparison. The two benchmark sources differ in how their model runs and outputs are obtained.

B.2 PREVIOUSLY REPORTED RESULTS

  • OSWorld 2.0 comparisons include published 500-step results, retaining reported scores and estimated model cost per task.Models use official screenshot implementations, with Claude models on their native computer-use interface and others generating PyAutoGUI actions.Source: §B.2 · p. 16
  • For BrowseComp-Plus, GPT-5 and o3 reference results use benchmark authors’ BM25 runs, with available final responses graded using Qwen3-32B and the official prompt.Source: §B.2 · p. 16
  • The GPT-OSS-120B BrowseComp-Plus reference uses the BM25 configuration as its sparse retriever.Source: §B.2 · p. 16

B.3 EVALUATION AND RESOURCE ACCOUNTING

Resource totals aggregate recorded agent and context-management usage, with token and cost totals calculated from calls and averaged across reported task sets.

B.3 EVALUATION AND RESOURCE ACCOUNTING

  • Token totals include recorded agent and context-management calls, with input processed again in later requests counted again, including cached input.Reported totals include Tracker and Compiler calls for CarryOn and summarizer calls for summarization-based methods; evaluator, user-simulator, and infrastructure costs are excluded.Source: §B.3 · pp. 16–17
  • Each call’s token total sums uncached input, cache-read, cache-write, and output tokens, counting reasoning already included in output only once.Cache-write lifetime breakdowns partition the overall cache-write count rather than adding to it.Source: §B.3 · p. 16
  • Nominal API-equivalent cost sums token counts multiplied by fixed category rates across calls, using the same rates for the agent and CarryOn’s auxiliary models.Cached input remains included at its corresponding cache price, so token-usage and cost reductions can differ.Source: §B.3 · p. 16
  • Per-method task totals are averaged over the reported task set, and relative token-usage and cost reductions are computed from those means.Calculations use unrounded values, with rounding applied only to displayed results.Source: §B.3 · p. 17

C.1 STATISTICAL PROCEDURES FOR CONTEXT AND CARRIER ANALYSIS

The section specifies how context and cost dynamics, post-fold text length, and carrier continuity are measured across trajectories and folds.

C.1 STATISTICAL PROCEDURES FOR CONTEXT AND CARRIER ANALYSIS

  • Context and cost dynamics use per-call traces across trajectories, defining each turn’s active set as trajectories with at least that many agent responses.Input counts include uncached input, cache reads, and cache creation, with screenshot tokens included; cost accumulates agent costs and runtime costs assigned to the same turn boundary.Source: §C.1 · p. 17
  • Figure 3 smooths per-turn means for visualization, while annotations use raw data and cost-curve shading shows their difference.The curves use Gaussian smoothing with a 12-turn standard deviation; gray bars show the fraction of baseline trajectories still active.Source: §C.1 · p. 17
  • Post-fold text length measures textual context immediately after each installed fold, averaging first within tasks and then equally across tasks with at least one fold.It uses a shared text projection, excludes image data and fixed prompts, and is paired with task performance to compare retained text and performance.Source: §C.1 · p. 17
  • 80% of carrier observations were retained or updated across 494 committed and installed folds: 650 of 813 observations.The pooled observations comprise 278 retained, 372 updated, and 163 added records; these percentages describe record continuity across folds.Source: §C.1 · p. 17
  • Table 9 presents selected steps from a main-experiment trajectory continuing from a populated form, with submission confirmation visible in the saved screenshot.Entries are condensed from the recorded trajectory.Source: Table 9 · p. 18

C.2 CONTINUATION CASES AND ABLATION BEHAVIORS

The continuation cases illustrate how retained task progress can be reused after a fold, while the prompt instructions define how CarryOn identifies, maintains, and complements reusable carriers.

C.2 CONTINUATION CASES AND ABLATION BEHAVIORS

  • The populated-form case shows the agent checking existing entries and submitting without re-entering authors; a saved screenshot confirms submission.The paper subsequently appears in the Research Output list.Source: §C.2 · p. 18
  • With carrier records masked, the prize-task agent re-extracts a PDF and reopens a catalogue to reconstruct a price mapping it had already established and used in a plan.The full CarryOn run continues filling the form after the fold.Source: §C.2 · p. 18, Table 10 · p. 19

D PROMPTS

  • The prompt distinguishes information products, whose semantic content can be used away from the current process, from situated state, which requires returning to the same environment or session.External objects hold independently addressable content; information products are structured, evidence-linked claims compiled from interactions.Source: D PROMPTS · p. 20
  • The semantic carry rule retains a candidate version when deleting it would require meaningful rereading, rechecking, re-decision, retry, or reconstruction.It favors useful context per retained token and targets the active candidate rather than duplicating its parent.Source: D PROMPTS · p. 20
  • The prompt requires open obligations to be recorded separately from hidden continuity state, with evidence steps for obligations shown as completed in the live tail.Obligations include required deliverables not completed, unperformed instruction steps, and values entered differently from their sources.Source: D PROMPTS · pp. 20, 22
  • Execution Context complements carriers with current goals and continuation information, while preserving requirements that a related work product may not satisfy.It tracks unfinished work and dependencies, and distinguishes completed implementation from pending testing or attempted submission from confirmed results.Source: D PROMPTS · p. 23

Related papers

Loading 2610.18784v2…