AstronOS：面向长视野智能体系统的统一执行模型与运行时
文章背景与核心概要
当前的智能体（Agent）系统通常将执行状态绑定到单一的对话、模型调用或孤立的智能体实例中，即便面对跨越多个阶段和长时间跨度的复杂任务时也是如此。为了解决这一痛点，科大讯飞等机构的研究人员推出了 AstronOS。它引入了一种统一的执行模型和运行时，旨在跨调用保持持久的工作项、唯一标识符以及带版本的权威状态。
AstronOS 通过将输入限定于特定的状态版本，并在经过严格验证后才推进状态，显著提升了多阶段执行管道的可靠性，超越了传统的上下文传递方法（如全历史回放或朴素摘要）。实验表明，在多阶段基准测试中，AstronOS 实现了极高的端到端通过率，为长视野智能体系统的工程落地提供了一条可靠的新途径。
执行摘要 (Executive Summary)
Current agentic systems often tie their execution states to a single conversation, model invocation, or isolated agent instance—even when complex tasks span multiple stages and calls over a long horizon. AstronOS introduces a unified execution model and runtime designed to maintain persistent work items, unique identities, and versioned authoritative states across calls.
Current agentic systems often tie their execution states to a single conversation, model invocation, or isolated agent instance—even when complex tasks span multiple stages and calls over a long horizon. AstronOS introduces a unified execution model and runtime designed to maintain persistent work items, unique identities, and versioned authoritative states across calls.
By scoping inputs to specific state versions and advancing states only after strict validation, AstronOS significantly improves reliability in multi-stage execution pipelines compared to traditional context-passing methods (such as full-history replay or naive summarization).
By scoping inputs to specific state versions and advancing states only after strict validation, AstronOS significantly improves reliability in multi-stage execution pipelines compared to traditional context-passing methods (such as full-history replay or naive summarization).
论文元数据 (Paper Metadata)
- arXiv ID:
arXiv:2608.16381[cs.AI] - Submission Date: August 17, 2026
- Primary Subject: Artificial Intelligence (
cs.AI) - Authors:
- Zhenhang Nie (1)*
- Gui Zheng (1) (Corresponding Author)*
- Xudong Sun (1)
- Tailong Zhu (1)
- Bin Zhang (1)** (Corresponding Author)
- (1) iFLYTEK Co., Ltd., Hefei, China
- Full-Text Links: View PDF | HTML Version | TeX Source
- arXiv ID:
arXiv:2608.16381[cs.AI]- Submission Date: August 17, 2026
- Primary Subject: Artificial Intelligence (
cs.AI)- Authors:
- Zhenhang Nie (1)*
- Gui Zheng (1) (Corresponding Author)*
- Xudong Sun (1)
- Tailong Zhu (1)
- Bin Zhang (1)** (Corresponding Author)
- (1) iFLYTEK Co., Ltd., Hefei, China
- Full-Text Links: View PDF | HTML Version | TeX Source
摘要 (Abstract)
Agentic systems often organize execution and state around a single conversation, model invocation, or agent instance, even when real work spans many calls and stages. We introduce a unified execution model that maintains a work item's persistent identity and versioned authoritative state across calls.
Agentic systems often organize execution and state around a single conversation, model invocation, or agent instance, even when real work spans many calls and stages. We introduce a unified execution model that maintains a work item's persistent identity and versioned authoritative state across calls.
Each step receives input scoped to a specific state version and new material; a result advances state only after validation and recording. We implement selected paths of this model in AstronOS using Cases, Tasks, and Scenario Packs across central and local execution.
Each step receives input scoped to a specific state version and new material; a result advances state only after validation and recording. We implement selected paths of this model in AstronOS using Cases, Tasks, and Scenario Packs across central and local execution.
We compare five complete strategies for carrying an established software-version update plan into a fresh model session: 1. Rereading original materials 2. Replaying full history 3. Deterministic text summary 4. Deterministic JSON 5. The AstronOS runtime-mediated handoff
We compare five complete strategies for carrying an established software-version update plan into a fresh model session: 1. Rereading original materials 2. Replaying full history 3. Deterministic text summary 4. Deterministic JSON 5. The AstronOS runtime-mediated handoff
实验设置与结果 (Experimental Setup & Results)
Ten controlled tasks were run under all five strategies with three repetitions, yielding 150 included executions:
Ten controlled tasks were run under all five strategies with three repetitions, yielding 150 included executions:
- Single-Stage Reference Family: All strategies performed similarly.
- Primary Three-Stage A-C Batch:
- AstronOS passed the frozen scorer in 14 of 15 executions.
- Rereading original materials passed 0 of 15 executions.
- Full-history replay passed 2 of 15 executions.
- Non-interleaved text summary and JSON batches each passed 0 of 15 executions.
- Cost vs. Time: AstronOS achieved a lower attempt-accounted model-token cost per passing execution, though it required more execution-window time per attempt.
- Single-Stage Reference Family: All strategies performed similarly.
- Primary Three-Stage A-C Batch:
- AstronOS passed the frozen scorer in 14 of 15 executions.
- Rereading original materials passed 0 of 15 executions.
- Full-history replay passed 2 of 15 executions.
- Non-interleaved text summary and JSON batches each passed 0 of 15 executions.
- Cost vs. Time: AstronOS achieved a lower attempt-accounted model-token cost per passing execution, though it required more execution-window time per attempt.
These findings associate the complete AstronOS condition with significantly higher end-to-end pass rates across fresh sessions in this benchmark, achieved at a measurable time cost.
These findings associate the complete AstronOS condition with significantly higher end-to-end pass rates across fresh sessions in this benchmark, achieved at a measurable time cost.