Long-horizon agents repeatedly reprocess interaction histories to continue work, wasting tokens and reconstruction effort. The paper's core insight is that continuation can be organized around reusable progress (task objects) rather than the histories that produced it. By exposing durable work products and explicit reuse conditions, an agent can inspect and reuse progress directly while still retaining complementary execution cues.
Key Findings
- Object-oriented context: The runtime separates context into continuation carriers (reusable objects and conditions for reuse), execution context (active goals, constraints, unresolved issues), and a live tail (recent interactions). This division lets the agent retire past history while keeping what is actually needed for continuation.
- Efficient folding and validation: CarryOn maintains carrier records across folds and revises only affected records by binding supporting evidence and running environment checks, avoiding full reconstruction each fold — so what: lowers repeated reconstruction cost and preserves validity of retained objects.
- Empirical gains: On OSWorld2.0, CarryOn improves success rate by 5.8% and cuts total token usage by 51.9% (cost −35.9%); on BrowseComp-Plus it improves success by 7.0% and reduces tokens by 22.9% (cost −7.1%). These numbers show both performance and resource savings on long-horizon computer-use and browsing tasks.
Who it's for and tradeoffs
Great fit if you develop or evaluate autonomous agents that interact across long sessions and need to preserve intermediate artifacts (files, session state, findings) without replaying full histories. CarryOn is especially useful when environment state or external objects can be re-opened or rechecked, and when token/cost efficiency matters. Look elsewhere if your tasks are short-lived, fully stateless, or cannot be expressed as reusable objects with explicit reuse conditions — the overhead of tracking carriers and grounding checks may not pay off for very small contexts.
Where it fits
CarryOn complements generic history compression and summarization: compressing history alone is insufficient for continuation when reusable progress exists. The paper positions CarryOn as a runtime strategy for long-horizon agent systems that balances correctness (via evidence binding and environment checks) with efficiency (fewer tokens, fewer reconstructions).