UCSD's CarryOn boosts agent OSWorld score 43.9→49.7 while cutting tokens 52%
mark_k · x · 2026-10-20
A UC San Diego paper introduces CarryOn, an object-oriented context runtime that lets long-horizon agents continue from completed work instead of repeatedly rereading how that work was produced.
Core idea
- Organizes continuation around reusable task objects (work products, grounded findings, usable app state) rather than compressed histories
- Three components: continuation carriers (bind object references and conditions of use), execution context (active goals and constraints), and a live tail (interactions since the last successful fold)
- Records are maintained across folds via bound evidence and environment checks, avoiding full representation rebuilds
Results (same Claude Opus 4.7 model)
- OSWorld 2.0 task-progress score: 43.9 → 49.7 (+5.8%); BrowseComPlus +7.0%
- Cost per task fell from $36.83 to $23.62; total token usage down 51.9% / 22.9%
Takeaway: better context management beats piling on tokens—performance and cost can improve together.
Related event: UCSD's CarryOn Runtime Cuts Agent Token Use by 52% While Boosting Accuracy(2 posts)→
More from coding & agent
- EvoWeb benchmark: 320 sequential requests expose coding agents' cross-request integration failures — TD-Labs · 2026-10-20
- Amazon's Lean4AgenticRL grounds multi-level process rewards in Lean4 for agentic RL — amazon · 2026-10-20
- Spectral Labs' AI agent passes SOLIDWORKS Professional CAD certification in under 3 hours — ninamiolane · 2026-10-20
- Dev shares single-session subagent workflow: parallel worktrees, HTML before/after reviews — AccBalanced · 2026-10-20
- Ask Reddit: Which non-coding agent workflows actually stuck in production — Live_Hedgehog8785 · 2026-10-20
- Using Codex to find cute deer in wildlife cam footage and cut a mini video — sch · 2026-10-20