Carry the agent's result, not its history

Long-running agents keep re-reading how they got somewhere. A new paper says to carry the result instead of the history.

The paper, from UC San Diego (arXiv, 14 October), introduces CarryOn. Say an agent has already put images into a slide deck. The deck itself shows that progress, but the agent keeps processing a growing log of the steps that produced it. That's wasted tokens.

CarryOn builds the agent's context around reusable task objects plus a short live tail of recent actions, instead of a compressed history. When something changes, it revises only the affected records and carries the rest forward.

Reported results: on OSWorld2.0, a 5.8% improvement with 51.9% fewer total tokens. On BrowseComp-Plus, a 7.0% improvement with 22.9% fewer tokens.

Better and cheaper at once is rare, so it's worth watching. It's still one paper on two benchmarks, so don't expect those exact numbers on your agent.

The idea you can use now: keep the agent's state as structured objects (the file, the table, the open tickets) and show the model that plus the last few steps, not the whole chat log.

How does your agent carry context across a long task: full history, a summary, or structured state?