SWE-AGILE: A Dynamic Reasoning Framework to Solve the Context Explosion Problem for AI Programming Agents

Addressing the context management dilemma of AI programming agents in software engineering tasks, SWE-AGILE proposes a two-layer dynamic reasoning strategy combining sliding window and reasoning summarization, setting a new record on SWE-Bench-Verified with 7B-8B parameter models, balancing reasoning depth and context efficiency.

In recent years, AI programming agents have shown significant potential, but they face context management challenges in complex tasks: traditional ReAct methods lack deep reasoning capabilities; when reasoning models extend Chain-of-Thought (CoT), they face a dilemma—retaining full history leads to context inflation (Lost-in-the-Middle problem), while discarding history results in repeated reasoning and wasted computation. This dilemma is particularly prominent in the SWE-Bench benchmark.

Two-Layer Context Architecture

• Sliding Window: A fixed-size buffer that stores recent complete reasoning to ensure immediate continuity
• Reasoning Summarization: Compresses historical reasoning into key conclusions, preserving core value

Dynamic Balance Mechanism

Adaptively adjusts window size and summary granularity based on task phases (exploration/convergence/backtracking)

Technical Details

• Summary Generation: Rule extraction, learning-based compression, hybrid strategies
• Sliding Window Management: Selects content based on importance and updates summaries incrementally

Achievements on the SWE-Bench-Verified benchmark:

• Scale Efficiency: 7B-8B models set a new performance standard (previous leading methods relied on 70B+ models)
• Data Efficiency: Trained with only 2.2k trajectories + 896 tasks
• Cost-Effectiveness: Significant reduction in reasoning costs Comparative advantages: More consistent reasoning quality, higher computational efficiency, stronger scalability

Implications

1. Reasoning depth and efficiency can be achieved simultaneously
2. Context is a scarce resource that requires careful management
3. The potential of small models is underestimated

Application Scenarios

Automated code review, intelligent debugging assistants, legacy code modernization, development tool integration

Limitations

• Risk of information loss in summaries
• The strategy is optimized for software engineering; cross-domain migration requires adjustments
• Reduced interpretability of the decision-making process

Future Directions

• Adaptive summary generation
• Hierarchical context management
• Cross-domain application expansion
• Context sharing for human-AI collaboration

SWE-AGILE solves the contradiction between deep reasoning and efficiency through dynamic context management, demonstrating the value of architectural innovation. The research team has open-sourced the code, providing an important reference for the fields of AI programming and agent architecture, and its design ideas are expected to be widely applied in future tools.