CarryOn · project page▶ watch the overview

Pack light,
CarryOn.

Keeping AI agents on track in long-horizon tasks

1–18 37–54 55–72 73–90 91–108 109–126 127–144 145–162 181–198 19–36 fold the whole history packed light

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

Yuchen Lu1,  Julian McAuley1,  Lei Zhang1,*

1University of California San Diego

*corresponding author

Paper arXiv Code Video

a two-minute overview

1 / 4 · The idea

Let task objects carry the progress.

A long task leaves its progress in the environment, yet the agent keeps re-reading how it got there. CarryOn builds the context around reusable task objects instead, so the agent carries just three things.

1Where the work is

Continuation carriers point to task objects, such as a file or an app state.

2What's left to do

Execution context keeps goals, constraints and open issues.

3Latest steps

The live tail keeps the steps since the last fold.

Two OSWorld 2.0 tasks before and after a fold

Folding with task objects on OSWorld 2.0. The presentation and the report carry the finished edits, while remaining constraints and calculations stay explicit. Active context drops from 42.4K to 14.4K tokens (−66.0%) and from 58.4K to 23.5K tokens (−59.7%).

2 / 4 · How it works

CarryOn runs alongside the agent.

The agent keeps its usual loop and reads a compact context view. Every so often, CarryOn tracks reusable progress, checks it against the environment, and commits one validated fold.

Overview of the CarryOn framework

3 / 4 · Results

Higher scores, with about half the tokens.

same agent model (Claude Opus 4.7), with and without CarryOn

OSWorld 2.0

long, real computer tasks · partial score (%)

43.9
49.7
without
with CarryOn
+5.8 points

tokens −52%40.1M → 19.3M per task

cost −36%$36.83 → $23.62 per task

BrowseComp-Plus

long web research · accuracy (%)

49
56
without
with CarryOn
+7 points

tokens −23%1.56M → 1.20M per task

cost −7%$1.89 → $1.76 per task

Task performance on OSWorld 2.0

Task performance on OSWorld 2.0. * Reproduced with the official benchmark and API.

4 / 4 · Long tasks

The context stays small, even on very long tasks.

Without CarryOn, the agent's context keeps growing with every turn. With it, the context levels off, so the longer the task runs, the more it saves.

Context size and cumulative cost over agent turns

Input context stays near 65K tokens while the baseline keeps growing; cost is 34.1% lower by turn 500.

BibTeX

@article{lu2026carryon,
  title   = {Pack Light, CarryOn: Building Object-Oriented Context for Long-Horizon Agents},
  author  = {Lu, Yuchen and McAuley, Julian and Zhang, Lei},
  journal = {arXiv preprint arXiv:2610.18784},
  year    = {2026}
}

An agent doesn't need to remember every step.

Just where the work is,
and what's left to do.

packed light!