Yuchen Lu, Julian McAuley and Lei Zhang build an agent’s working context around reusable task objects. A saved document can hold completed edits. A separate record keeps its location, supporting evidence and conditions for reuse, while unfinished requirements remain explicit.
Their CarryOn runtime keeps three kinds of context:
- Records of reusable progress: files, source-grounded findings or application state.
- Current goals, constraints and open obligations that those objects do not capture.
- Recent interactions since the last accepted context update.
The runtime checks evidence and record consistency before replacing older interactions in the active prompt. Failed updates preserve the previous context. Those checks establish what can be carried forward; they do not verify that the task is correct.
With the same Claude Opus 4.7 agent on OSWorld 2.0 computer tasks, the authors report:
- Partial credit rises from 43.9% to 49.7%; fully completed tasks rise from 15.7% to 19.4%.
- Total tokens per task fall from 40.10 million to 19.30 million, including context-management calls.
- Nominal model cost falls from $36.83 to $23.62 per task. On a separate 100-question research test, accuracy rises from 49% to 56%, but cost falls only 7.1%.
A shared 40-task comparison also favors CarryOn over generic summaries on partial credit, while matching the uncompressed baseline’s full-completion rate. Saving tokens helps only if the agent retains both accomplished work and what it still owes.
Caveat: these are author-reported results with one host model. The repository still promises implementation code after review, so the runtime is not yet a released package to reproduce them.