PaperlayerView exact revision on arXiv

Source-linked AI summary

openJiuwen: Beyond Static Harnesses for Long-Horizon Coding Agents

openJiuwen Team, Tao Yu, Xinyu Zhang, Qianqian Chen, Xiaoneng Xiang, Chia Kwangyang, Xingchen Huang, Ran Chen, Yangkai Ding, Zheng Wang, Yeo Boon Hong, Bingzheng Gan, Enrui Hu, Shuo Cheng, Deyang Li, Ruifeng Shi, Hongbo Wang, Qi Ye, Xuefeng Jin, Zhangchun Zhao

arXiv:2608.27969v1cs.AI

TL;DR

Long-horizon coding agents need harnesses that compose heterogeneous capabilities and adapt execution as task evidence evolves. openJiuwen combines shared execution semantics and Rail-based composition with adaptive framework-controlled runtime decisions; it exceeds the strongest selected leaderboard results across SWE-bench Verified and Terminal-Bench 2.1 by 3.4 and 3.39 percentage points, respectively.

  • Problem

    Long-horizon coding agents must maintain coherent behavior as repository state, diagnostics, task progress, and relevant context evolve, while their harness assembles capabilities and controls execution.

    Source: §1 · pp. 1–2

  • Method

    openJiuwen uses a shared execution substrate and Rail-based capability composition across single agents, delegated sub-agents, and Swarm Flow, while adapting framework-controlled runtime state around a fixed model policy.

    Source: §1 · p. 3

  • Results

    Across both benchmarks, openJiuwen exceeded the strongest selected leaderboard results by 3.39 percentage points on Terminal-Bench 2.1 and 3.4 percentage points on SWE-bench Verified.

    Source: §4.3 · p. 16

  • Takeaways & Limitations

    Consistent strong performance across repository-level issue resolution and terminal-based execution supports openJiuwen as a reusable harness for heterogeneous long-horizon workloads.

    Source: §4.3 · p. 16

    The evaluation focuses on SWE-bench Verified and Terminal-Bench 2.1; broader controlled studies across additional benchmarks, models, and configurations, plus more detailed ablations, remain future work.

    Source: §5 · p. 16

Abstract

from arXiv · show

Long-horizon coding agents operate over evolving repository states while increasingly relying on heterogeneous capabilities, delegated agents, and multi-agent coordination. These trends pose two complementary challenges for the agent harness. First, developers need to compose capabilities, reconfigure execution logic, and scale increasingly complex agent systems without repeatedly rebuilding orchestration. Second, complex coding tasks continuously produce new evidence---such as semantic diagnostics, execution outcomes, task progress, and changing context relevance---that should dynamically influence subsequent runtime decisions. We characterize these challenges as Structural Composability and Runtime Adaptivity. We present openJiuwen, an open-source harness designed for both developer composability and adaptive task execution. openJiuwen provides a shared execution substrate and Rail-based capability composition across single agents, delegated sub-agents, and Swarm Flow, enabling developers to construct sophisticated agent harnesses under common execution semantics. It further adapts framework-controlled runtime decisions around a fixed model policy, allowing evolving evidence to dynamically affect context, feedback, and task control toward successful completion. We systematically evaluate openJiuwen on SWE-bench Verified and Terminal-Bench 2.1, where it achieves 82.6% and 87.19%, respectively, exceeding the strongest selected official-leaderboard point estimates by 3.4 and 3.39 percentage points. These results show that openJiuwen achieves strong performance on complex coding tasks while providing a composable and adaptive harness design.

1. Introduction

Long-horizon coding agents must coordinate capabilities while adapting execution as repository state and task evidence evolve. openJiuwen addresses these challenges with a shared, composable harness that adapts framework-controlled decisions around a fixed model policy.

1. Introduction

  • Long-horizon coding agents must maintain coherent behavior as code state, diagnostics, task progress, and relevant context evolve during execution.Planning, memory, verification, context engineering, delegation, and multi-agent coordination make the harness a systems layer for assembling capabilities and controlling execution.Source: §1 · pp. 1–2
  • Structural Composability uses a shared execution substrate and Rail-based capability composition across single agents, delegated sub-agents, and Swarm Flow.Nested Inner Loop/Outer Loop control and lifecycle hooks, ordered composition, and visibility gating reuse the same execution semantics, reducing topology-specific orchestration.Source: §1 · pp. 2–3
  • Runtime Adaptivity changes framework-controlled execution state around a fixed model policy so emerging evidence can influence context, feedback, continuation, and stopping.openJiuwen implements this through Context Management, Goal Mode, LSP-Driven Passive Feedback, and Self-Reflection, which reuses distilled experience across tasks.Source: §1 · p. 3
  • 82.6% on SWE-bench Verified and 87.19% on Terminal-Bench 2.1 exceed the strongest selected official-leaderboard results by 3.4 and 3.39 percentage points, respectively.Further SWE-bench Verified analysis finds openJiuwen remains competitive as task duration increases.Source: §1 · p. 3

2. Related Work

Related work advances coding-agent scaffolds, runtime controls, and context management through specialized mechanisms. openJiuwen positions the harness as a shared substrate for composing capabilities and adapting execution to evolving feedback, progress, and context.

2.1. Coding-Agent Architectures and Scaffolds

  • Coding-agent systems improve repository interaction, action spaces, environments, and specialized capabilities, while reusable harnesses for shared execution across single and multi-agent settings receive less attention.Examples include SWE-agent Yang et al. [2024], CodeAct and OpenHands Wang et al. [2024, 2025], AutoCodeRover Zhang et al. [2024], LingmaAgent Ma et al. [2025], RepairAgent Bouzenia et al. [2025], Agentless Xia et al. [2025a], Trae Agent Trae Research Team et al. [2025], and multi-agent Swarm Skills Zhang et al. [2026].Source: §2.1 · pp. 3–4

2.2. Runtime Behavior for Long-Horizon Coding Agents

  • Long-horizon runtime work addresses repeated actions through execution-state tracking and workflow optimization, but evolving semantic feedback, task progress, and context relevance motivate broader runtime adaptation.Studies report recurring action patterns and longer, more variable failed trajectories despite file localization Lindenbauer et al. [2025]; Majgaonkar et al. [2025]; Ledger Wang et al. [2026b] tracks execution state, while Collaborative Optimization Zhang et al. [2025a] adapts workflows, prompts, demonstrations, and routing.Source: §2.2 · p. 4

2.3. Context Management for Long-Horizon Agents

  • Context-management and prompt-adaptation methods change model-visible inputs, while openJiuwen treats the harness as the layer where feedback, progress, and context changes shape subsequent execution.SWE-AGILE Lian et al. [2026], SWE-Pruner Wang et al. [2026a], AgentDiet Xiao et al. [2026], observation masking Bouzenia and Pradel [2025], and P3 Zhang et al. [2025b] adapt visible context or prompts; openJiuwen also composes and scales agent capabilities through a shared execution substrate.Source: §2.3 · p. 4

3. Method: Composable Structure and Adaptive Runtime Control

openJiuwen separates structural composability, which organizes capabilities and agent structures over a shared execution substrate, from runtime adaptivity, which changes execution as new information emerges.

Runtime adaptivity

  • Runtime adaptivity changes context, semantic feedback, task progress, and reusable experience as new information emerges.The illustrated mechanisms include Context Management, Goal Mode, LSP-Driven Passive Feedback, and Self-Reflection; they are representative rather than exhaustive.Source: §3 · p. 4, Figure 2 · p. 5

Structural composability

  • Structural composability combines Rail-based capability composition with reuse from single agents and delegated sub-agents to Swarm Flow over a shared execution core.The shared core comprises Inner Loop and Outer Loop execution.Source: Figure 2 · p. 5

3.1. Structural Composability

openJiuwen composes standalone agents, delegated sub-agents, and Swarm Flow on a shared execution substrate, using layered loops and ordered Rails to add capabilities without changing the core engine. Swarm Flow lets developers combine coordination operators into task-specific multi-agent workflows.

3.1. Structural Composability

  • Standalone agents, sub-agents, and Swarm Flow reuse the same execution engine and Rail interface, so scaling changes composition rather than the underlying execution substrate.This shared substrate supports distinct capability configurations across participating agents.Source: §3.1 · p. 5, §3.1.2 · p. 7

3.1.1. Inner Loop and Outer Loop: Layered Execution

  • The Inner Loop manages bounded model–tool interaction, while the Outer Loop evaluates task-level continuation after each complete Inner Loop round.Composable evaluators can use semantic completion signals, resource limits, or user-defined conditions; the Outer Loop also supports bounded retry or recovery.Source: §3.1.1 · pp. 5–6

3.1.2. Rail: Ordered Capability Composition

  • Rails attach cross-cutting capabilities to lifecycle hooks and execute by declared priority, while visibility gating enables role-specific capability configurations without changing the loops.Capabilities can include security, memory, planning, context engineering, semantic feedback, delegation, and human interaction; priority ordering is deterministic for ties.Source: §3.1.2 · pp. 6–7

3.1.3. Swarm Flow: Composable Multi-Agent Coordination

  • Swarm Flow combines composable operators for budgets, parallel branches, result filtering and streaming, stateful sessions, human intervention, and termination into task-specific workflows.Developers can rearrange, omit, and combine operators rather than adopt a fixed multi-agent architecture.Source: §3.1.3 · pp. 7–8

3.2. Runtime Adaptivity

openJiuwen adapts framework-controlled context, feedback, and stopping behavior as evidence changes during execution, and carries selected experience into future tasks, all around a fixed model policy.

3.2.1. Constrained Online Runtime Adaptation

  • Runtime adaptivity treats within-task execution as constrained, state-dependent adjustment of context, semantic feedback, and stopping under fixed policy π, rather than learned model changes.Interaction history, diagnostics, goal state, context pressure, and resource status inform adjustments; the conceptual trajectory objective weighs task utility against runtime cost, with stopping governed by acceptance and resource conditions.Source: §3.2.1 · pp. 8, 10, §3.2 · p. 8

3.2.2. Context Management: Adaptive Context Construction

  • Context Management adapts model inputs to context pressure, task state, and content structure through progressive compression, structure-aware reduction, dead-loop collapse, and artifact offloading with coarse-to-fine retrieval.Recent information is preferentially retained at higher fidelity, and the assembled context remains subject to a budget constraint.Source: §3.2.2 · p. 10

3.2.3. Goal Mode: Acceptance-Constrained Stopping

  • Goal Mode maintains an explicit task goal and separates semantic completion or blockage from configured attempt, time, or resource limits when deciding whether execution should continue.It supports self-assessment, independent assessment, and hybrid assessment through one interface.Source: §3.2.3 · pp. 10–11

3.2.4. LSP-Driven Passive Feedback: Closed-Loop Semantic Correction

  • LSP-driven passive feedback automatically surfaces mechanically produced, filtered diagnostics after relevant code changes, while active semantic queries are explicitly requested by the agent.Its scope is limited to properties exposed by the analysis infrastructure, such as type errors and static diagnostics; it does not establish architectural quality, maintainability, business-logic correctness, or task success.Source: §3.2.4 · p. 11

3.2.5. Self-Reflection: Cross-Task Non-Parametric Adaptation

  • Self-Reflection extracts and stores reusable experience after a trajectory, then retrieves relevant experience for later tasks without updating model policy π.Retrieved experience remains subject to Context Management and may be compressed, summarized, deferred, or omitted when constructing the later task’s context.Source: §3.2.5 · pp. 11–12

4. Experiments

Experiments evaluate openJiuwen on repository issue resolution and containerized terminal tasks, finding strong results on both benchmarks and across longer execution horizons.

4.1. Experimental Setup

  • The setup covers 500 SWE-bench Verified repository tasks and 89 Terminal-Bench 2.1 terminal tasks, reporting Pass@1 and verifier-based Accuracy, respectively.SWE-bench uses Claude Opus 4.5 with high reasoning effort; the primary Terminal-Bench result uses GPT-5.6 Sol, with Fable 5 additionally used for model-matched comparisons. Prompt, tools, and model configuration are fixed within each openJiuwen configuration.Source: §4.1 · p. 12

4.2. Results on Terminal-Bench 2.1

  • 87.19% accuracy on Terminal-Bench 2.1 beats Claude Code with Fable 5 at 83.8% by 3.39 points; 82.6% Pass@1 on SWE-bench Verified beats 79.2% by 3.4 points.The cross-benchmark comparison is at the system level, not a controlled estimate of the harness contribution alone.Source: Table 1 · p. 13, §4.2 · p. 13, Table 2 · p. 13
  • With Fable 5, openJiuwen scores 0.76 on file operations (vs. 0.56 for Claude Code and 0.52 for Terminus 2) and 0.889 on system administration (vs. 0.778 and 0.844).GPT-5.6 Sol scores 0.956 on system administration, 0.908 on software engineering, and 0.950 on data science. The authors suggest operational tools and a unified execution interface may contribute, while noting the benchmark does not isolate tool availability causally.Source: Table 3 · p. 14, Figure 5 · p. 14, §4.2 · p. 15

4.3. Results on SWE-bench Verified

  • OpenJiuwen leads the < 15 minute (91.75%) and 15-minute–1-hour (81.23%) buckets and scores 52.38% on 1–4 hour tasks, versus 35.71% for mini-swe-agent with Opus 4.5 high.On 1–4 hour tasks, it is close to live-SWE-agent at 54.76%; the > 4 hour bucket has only three tasks, so the paper draws no strong conclusion there.Source: §4.3 · p. 15, Table 4 · p. 15
  • For mini-swe-agent on 1–4 hour tasks, Opus 4.5 high scores 35.71% versus 42.86% with medium reasoning effort; openJiuwen scores 52.38% with Opus 4.5 high.The paper interprets this as consistent with selective context management mitigating context pressure, but does not claim a controlled causal finding.Source: §4.3 · pp. 15–16

5. Limitations

The paper identifies several directions for further development of openJiuwen, spanning runtime mechanisms and evaluation scope.

5. Limitations

  • The authors propose tighter integration of runtime Self-Reflection and offline evolution to reuse execution experience more systematically.The current design treats them as complementary stages.Source: §5 · p. 16
  • Goal Mode currently targets task-level objectives; extending it to hierarchical goals may improve control over complex, multi-stage tasks.Source: §5 · p. 16
  • The authors propose automating Context Management strategy selection and coordination under different runtime conditions.Source: §5 · p. 16
  • Evaluation is limited to SWE-bench Verified and Terminal-Bench 2.1; broader controlled studies and more detailed ablations could better characterize generality and mechanism contributions.The proposed studies would span additional benchmarks, models, and configurations.Source: §5 · p. 16

6. Conclusion

openJiuwen addresses Structural Composability and Runtime Adaptivity in coding-agent harnesses. It combines reusable execution semantics with evidence-guided runtime decisions and demonstrates strong performance across two coding-task benchmarks.

6. Conclusion

  • openJiuwen provides shared execution infrastructure and Rail-based composition across single agents, delegated sub-agents, and multi-agent coordination to reduce harness development effort.Its shared Inner Loop/Outer Loop execution substrate supports reusable execution semantics for constructing and extending sophisticated harnesses.Source: §6 · p. 16
  • openJiuwen adapts framework-controlled runtime decisions so evolving execution evidence guides subsequent behavior on complex coding tasks.Its mechanisms include Goal Mode, LSP-Driven Passive Feedback, Context Management, and Self-Reflection.Source: §6 · p. 16
  • Evaluations on SWE-bench Verified and Terminal-Bench 2.1 demonstrate strong performance across repository-level software engineering and broader terminal-based tasks.The conclusion presents composable harness architecture as supporting both efficient agent development and effective execution on complex coding tasks.Source: §6 · p. 16

Related papers

Loading 2608.27969v1…