AIAny
Icon for item

StateAct: Program State, before Pixels, for Long-Horizon Computer-Use Agents

Drives long‑horizon desktop agents by reading and manipulating program state (files, DOM, backends) instead of relying on screenshots. The main agent uses code for actions and structural verification while a lightweight GUI subagent handles rare screenshot-click steps, improving success rates and lowering per-task cost versus screenshot-only approaches.

Introduction

Most computer-use agents treat screenshots as the canonical observation—and pay for it: pixels hide the true program state (files, DOM, application backends), creating ambiguous perceptions and expensive perception bottlenecks. StateAct reverses that assumption by making program state the primary grounding for action, memory, and verification, pushing the system bottleneck from perception toward reasoning.

Key Findings
  • State-first orchestration: the main agent acts by executing and reasoning over program state (code-driven actions) rather than selecting pixel coordinates, with a dedicated GUI subagent reserved for 28 of 108 tasks and only 1.1% of main-agent steps. This reduces reliance on costly screenshot-based perception.
  • Measured gains on OSWorld 2.0: when driving Claude Opus 4.8, StateAct raises binary success from 20.6% to 26.9% and partial success from 54.8% to 61.6%, while delivering ~9× lower cost per task than a screenshot-driven pipeline. A code-only variant (no GUI subagent) achieves 45.9% partial success, below the screenshot baseline, showing the hybrid state+GUI balance matters.
  • Verification and robustness: an independent finish gate checks structural correctness (missing output, unsaved files, wrong paths), catching failures that pixel checks miss and enabling safer long-horizon plans.
  • Context management: to stay reliable over hundreds of steps, the main agent delegates subgoals to fresh subagents, keeping each agent's context compact and reducing compounded reasoning errors.
How it works

StateAct is a code-first, multi-agent harness. The main agent inspects and modifies program state via executable code or APIs; the GUI subagent performs screenshot-and-click interactions only when required. A finish-gate verifier reads saved results to ensure structural correctness rather than relying on visual checks. The architecture emphasizes state-grounding: actions, memory, and verification are anchored to machine-readable state rather than pixel observations.

Who it's for and tradeoffs

Great fit if you build or evaluate agents that operate on desktop/web applications and can access program state or APIs—especially long-horizon workflows where verification and low-cost repeated interactions matter. Look elsewhere if you target closed GUIs with no programmatic state access, purely vision-only benchmarks, or extremely resource-constrained environments where integrating program-state access and verifiers is infeasible. State-grounding improves correctness and cost but shifts failure modes toward reasoning and state-modeling mistakes rather than perception errors.

Information

  • Websitearxiv.org
  • AuthorsYan Yang, Xiangru Jian, Ziyang Luo, Zirui Zhao, Yutong Dai, Ziji Shi, Hanshu Yan, Jun Hao Liew, Silvio Savarese, Junnan Li
  • Published date2026/07/24

More Items

Bridges the proprietary-to-open-source gap in agentic search by converting multi-step retrieval and reasoning traces into a structured, style-normalized JSON protocol and using it for joint distillation + RL. Produces denser supervision that improves student success rates while reducing style drift.

Lets canvas-native agents plan, generate, edit, and organize long-horizon multimodal creative projects by representing artifacts, versions, and actions as typed canvas nodes and links. Uses a three-layer design (canvas state, protocol bridge, agent runtime) so agents act within an inspectable, editable project state.

Empirically studies how transformer-based native multimodal pre-training scales under fixed compute, deriving compute- and data-allocation power laws and an efficiency frontier for model size, token count, and data mixture; evaluates cross-modal transfer and multimodal in-context learning.