SWE-AGILE: A Software Agent Framework for Efficiently Managing Dynamic Reasoning Context

Shuquan Lian; Juncheng Liu; Yazhe Chen; Yuhong Chen; Hui Li

cs.AIApr 13, 2026

SWE-AGILE: A Software Agent Framework for Efficiently Managing Dynamic Reasoning Context

Authors: Shuquan Lian, Juncheng Liu, Yazhe Chen, Yuhong Chen, Hui Li

Organizations: Key Laboratory of Multimedia Trusted Perception and Efficient Computing Ministry of Education of China, Xiamen University · Microsoft

Abstract

Prior representative ReAct-style approaches in autonomous Software Engineering (SWE) typically lack the explicit System-2 reasoning required for deep analysis and handling complex edge cases. While recent reasoning models demonstrate the potential of extended Chain-of-Thought (CoT), applying them to the multi-turn SWE task creates a fundamental dilemma: retaining full reasoning history leads to context explosion and Lost-in-the-Middle'' degradation, while discarding it would force the agent to redundantly re-reason at every step. To address these challenges, we propose SWE-AGILE, a novel software agent framework designed to bridge the gap between reasoning depth, efficiency, and context constraints. SWE-AGILE introduces a Dynamic Reasoning Context strategy, maintaining a sliding window'' of detailed reasoning for immediate continuity to prevent redundant re-analyzing, while compressing historical reasoning content into concise Reasoning Digests. Empirically, SWE-AGILE sets a new standard for 7B-8B models on SWE-Bench-Verified using only 2.2k trajectories and 896 tasks. Code is available at https://github.com/KDEGroup/SWE-AGILE.

Figures & tables

Figure 1: A comparison of context growth patterns in the multi-turn SWE task: (1) Shallow Thinking maintains low context cost but lacks reasoning depth. (2) Interleaved Thinking enables deep System-2 reasoning but suffers from rapid, linear context growth (steep context bar). (3) SWE-AGILE enables a sustainable “Sawtooth” growth pattern. Within the Sliding Reasoning Window , the agent engages in deep reasoning (steep slope similar to Interleaved Thinking); however, as steps progress, historical thoughts are compressed into concise Reasoning Digests .

Figure 2: Overview of SWE-AGILE. (1) Trajectory Snapshot Training : We decompose long trajectories into discrete snapshots. In each snapshot, historical reasoning traces outside the Sliding Reasoning Window are replaced by digests ( $d_{i}$ ) and masked from the loss, forcing the model to learn the target $Y_{t}$ (Reasoning $r_{t}\to$ Digest $d_{t}\to$ Action $a_{t}$ ) based on a compact context $C_{t}$ . (2) Backfilling Reasoning and Digest : We synthesize high-quality training data by employing a reasoning model to augment raw trajectories. Conditioned on the ground-truth future action, the semantic hint and the dynamic context, the model backfills detailed reasoning ( $r_{t}$ ) and concise digest ( $d_{t}$ ). (3) RLVR with Trajectory Level Compression: We introduce a compression-aware reward function. The model is incentivized to maximize the reduction ratio between the actual hybrid context ( $\mathcal{L}_{hybrid}$ ) and the hypothetical full context ( $\mathcal{L}_{full}$ ), thereby learning to efficiently compress historical information without sacrificing task success ( $\mathbb{I}_{success}$ ).

Approach	Base Model	Scaffold	Data	Success Rate
Closed Weight Models
OpenHands ( Wang et al., 2025d )	GPT-5	OpenHands	-	71.80
OpenHands	Claude 4 Sonnet	OpenHands	-	70.40
SWE-agent ( Yang et al., 2024b )	Claude 4 Sonnet	SWE-agent	-	66.6
Agentless-1.5 ( Xia et al., 2025a )	GPT-4o	Agentless	-	38.8
SWE-RL ( Wei et al., 2025 )	Llama-3.3-70B	Agentless Mini	-	41.0

Table 1: Performance Comparison on SWE-Bench-Verified benchmark.

Inference Baselines (Base Model)
Method	Avg Steps	Success Rate (%)
(1) Disable Thinking	43.95	0.03
(2) Interleavd Thinking	26.08	12.42
(3) Current-Step Thinking	15.77	15.83
SFT
(4) Standard SFT (w/o Backfilling)	23.85	14.83

Table 2: Ablation study on context management and training strategies. We report the Average Steps per trajectory (as a proxy for exploration depth) and the Success Rate on SWE-bench-Verified. w/o CR denotes the ablation setting where the Compression Reward is excluded during RLVR.

Figure 3: Average Response Length per Step.

Appendix figures & tables4 assets

Supplementary material from the paper’s appendix.

Appendix

Figure 4: The System Prompt for SWE-AGILE. Coherent with Backfilling Prompt, it helps reducing redundant re-analyzing.

Figure 5: The Prompt for Backfilling Reasoning and Digest. Coherent with Agent System Prompt, it helps reducing redundant re-analyzing.

Hyperparameter	Value
Learning Rate	$1\times 10^{-6}$
Global Batch Size	16
Mini-Batch Size	8
Generations per Prompt ( $G$ )	8
Max Prompt Length	28,582
Max Response Length	4096

Appendix

Table 3: Hyperparameters for RLVR (DAPO) Training

Hyperparameter	Value
Learning Rate	$1\times 10^{-5}$
Batch Size	32
Max Sequence Length	26000 tokens
Total Epochs	4
Sliding Window Size	random in [2, 5]

Appendix

Table 4: Hyperparameters for Snapshots Training

Software Engineering Agents

Software Engineering Agents are AI systems that use code, tools, and feedback to perform software development tasks.

See all 344 papers

From SWE-ZERO to SWE-HERO: Execution-free to Execution-based Fine-tuning for Software Engineering Agents
Apr 2, 2026Nikolai Ludwig, Wasi Uddin Ahmad, Somshubra Majumdar +1Software Engineering Agents LLM Agent Training
NVIDIA Santa Clara, CA 95051, USA
SWE-Edit: Rethinking Code Editing for Efficient SWE-Agent
Apr 28, 2026Yikai Zhang, Jiaxin Pei, Kenan Li +9AI Coding Agents Multi-Agent Coding
1Microsoft, Redmond, Washington, United States · 3Stanford Institute for Human-Centered Artificial Intelligence (HAI), Stanford University · Department of Computer Sciences, University of Wisconsin–Madison
SWE-Router: Routing in Multi-turn Agentic Software Engineering Tasks
Jun 30, 2026Seongho Son, Sangwoong Yoon, Jiahua Tang +3Software Engineering Agents LLM Agent Workflow Optimization
University College London, United Kingdom · Ulsan National Institute of Science and Technology, South Korea · PSL Research University, France +1

Apr 2, 2026J/K move · Enter open · S save

From SWE-ZERO to SWE-HERO: Execution-free to Execution-based Fine-tuning for Software Engineering Agents

Nikolai Ludwig, Wasi Uddin Ahmad, Somshubra Majumdar, Boris Ginsburg

NVIDIA Santa Clara, CA 95051, USA

Software Engineering Agents LLM Agent Training Automated Software Engineering Fine-Tuning Automated Program Repair

Open paper AskPDF arXiv

We introduce SWE-ZERO to SWE-HERO, a two-stage SFT recipe that achieves state-of-the-art results on SWE-bench by distilling open-weight frontier LLMs. Our pipeline replaces resource-heavy dependencies with an evolutionary refinement strategy: (1) SWE-ZERO utilizes large-scale, execution-free trajectories to master code semantics and repository-level reasoning, and (2) SWE-HERO applies targeted, execution-backed refinement to transition these semantic intuitions into rigorous engineering workflows. Our empirical results set a new benchmark for open-source models of comparable size. We release a dataset of 300k SWE-ZERO and 13k SWE-HERO trajectories distilled from Qwen3-Coder-480B, alongside a suite of agents based on the Qwen2.5-Coder series. Notably, SWE-HERO-32B achieves a 62.2% resolution rate on SWE-bench Verified. Furthermore, despite being trained exclusively on Python, our agents demonstrate robust zero-shot transferability on SWE-bench Multilingual, reaching 44.1% and confirming the paradigm's generalizability across diverse languages.

Abstract

Software Engineering Agents

Explore similar work

From SWE-ZERO to SWE-HERO: Execution-free to Execution-based Fine-tuning for Software Engineering Agents

SWE-Edit: Rethinking Code Editing for Efficient SWE-Agent

SWE-Router: Routing in Multi-turn Agentic Software Engineering Tasks