CarryOn Framework Released: Revolutionizing Long-Horizon Agent Context Construction

Summary:CarryOn is a novel object-oriented context runtime framework designed to optimize long-horizon agent task state management. By reusing task objects instead of reprocessing interaction histories, CarryOn constructs a context view that maintains task state consistency while significantly reducing token usage. Experimental results on OSWorld 2.0 and BrowseComp-Plus benchmarks show performance improvements of 5.8% and 7.0%, respectively, and a reduction in token usage by 51.9% and 22.9%. This innova

---

Technical Mechanism Analysis

The core innovation of the CarryOn framework lies in its object-oriented context construction mechanism. Unlike traditional methods, CarryOn does not rely on compressed history records or reconstructed progress descriptions. Instead, it builds a context view through the following three key components:

1. Continuation Carriers: These are reusable objects that carry the key state information of the task.
2. Complementary Execution Context: Provides additional environmental information related to the current task.
3. Live Tail of Recent Interactions: Captures the latest interaction data to ensure the real-time and accuracy of the context view.

To address the challenge of maintaining the validity of the reusable basis after the surrounding history is retired, CarryOn adopts a method of maintaining carrier records across folds. It revises affected records against bound evidence and applicable environment checks while carrying forward unaffected records and jointly updating the execution context. This approach avoids the overhead of reconstructing the full representation at each fold, thereby improving efficiency.

Engineering Trade-offs and Experimental Performance

CarryOn balances the complexity of context construction with token usage. By reusing task objects, CarryOn reduces the need to reprocess history records, thereby lowering token usage. However, this also means that more complex management mechanisms are required to ensure the consistency and validity of task objects.

In the OSWorld 2.0 and BrowseComp-Plus benchmarks, CarryOn demonstrated significant advantages in performance improvement and token usage reduction. Specifically, CarryOn achieved a 5.8% performance improvement in OSWorld 2.0 while reducing token usage by 51.9%; in BrowseComp-Plus, the performance improved by 7.0% and token usage decreased by 22.9%. These results indicate that CarryOn is highly efficient and scalable in handling long-horizon agent tasks.

Developer Implementation and Deployment Recommendations

For developers, CarryOn provides a new approach to context management, particularly suitable for agent applications that need to handle long-horizon tasks and complex states. Here are some deployment recommendations:

1. Task Object Design: When designing task objects, fully consider their reusability and consistency to ensure the effectiveness of CarryOn.
2. Context Management Strategy: Develop a reasonable context management strategy according to the specific application scenario to balance performance improvement and resource consumption.
3. Performance Monitoring and Optimization: Continuously monitor the performance of CarryOn and optimize it according to actual needs to fully leverage its advantages.

Conclusion

The CarryOn framework provides an efficient object-oriented context construction mechanism for long-horizon agents. By reusing task objects and optimizing the context view, it significantly improves the efficiency and scalability of agents. This innovation provides AI developers with new tools and ideas, especially for handling complex tasks and long-horizon interactions.

---

Source: Hugging Face Daily Papers (2026-10-13)

— END —

Tags: #CarryOn Follow #Long-Horizon Agents Follow #Context Management Follow #Object-Oriented Follow #AI Framework Follow