← News

Agents & MCPSources: arXiv · 16 Nov 2026

CarryOn context manager pairs compact state with source-linked evidence for long-horizon LLM agents

A research paper introduces CarryOn, a lightweight context manager that keeps relevant evidence available to LLM agents without extra model calls, reporting higher task success than baselines across three benchmarks.

In plain words

AI agents that handle long, multi-step tasks accumulate long interaction histories, and deciding what to keep in context is hard: summaries can drop exact details, while full histories add clutter. CarryOn is a learned context manager that keeps a compact execution state plus fine-grained evidence linked to its source. In the paper's reported tests it raised task success while avoiding extra LLM calls for context management.

What was announced

The arXiv paper, published as an abstract, argues that existing context-reduction methods face a trade-off: shorter contexts do not necessarily support better decisions, because summaries can omit exact details needed for later actions, while retaining entire interactions carries substantial irrelevant information. CarryOn addresses this by pairing a compact execution state with fine-grained, source-linked evidence, using a learned selector and an adaptive retention policy. The authors state this keeps relevant evidence available without additional LLM calls for context management.

To evaluate the approach, the authors introduce two metrics: S-alpha, which prioritizes task success while accounting for resource use, and normalized cost per success (NCS), which measures resources spent per successful task. Across the AppWorld, OfficeBench, and Tau2 Telecom benchmarks with three LLM Actors, CarryOn reportedly achieved 75.7% macro-averaged task success, exceeding the strongest baseline by 6.7 percentage points, and the best average scores on both metrics. These figures come from the paper's own abstract and have not been independently verified.

What it does not show

  • The source is only the paper's abstract; the full article was not reviewed, so method details and evaluation conditions cannot be checked.
  • All performance figures are the authors' own reported results from the abstract; no independent confirmation is available.

Sources

Source-support review completed 17 Nov 2026. This story was drafted with AI assistance and checked against retained source passages. Automated checking is not independent verification.

Coverage: arXiv: abstract.

Updates and corrections

No corrections so far.

First published 17 Nov 2026

New to the subject? AI from the Ground Up starts from the beginning.