FastContext: Training Efficient Repository Explorer for Coding Agents

2026-06-18 · Source: cs.SE updates on arXiv.org · Field: Technology & Digital — Artificial Intelligence & Machine Learning, Software Development & Engineering · Depth: Expert, extended

Summary

FastContext is a novel exploration subagent designed to enhance Large Language Model (LLM) coding agents by decoupling repository exploration from task solving. This system, powered by specialized models ranging from 4B to 30B parameters, addresses the significant bottleneck of locating relevant code, which often consumes excessive token budgets and introduces irrelevant information into an agent's context. FastContext operates on demand, executing parallel tool calls (Read, Glob, Grep) to identify and return concise file paths and line ranges as focused context. Its models are trained using supervised fine-tuning from reference trajectories and further refined with reinforcement learning for broad initial search, multi-turn evidence gathering, and precise citation generation. Integrating FastContext into Mini-SWE-Agent has demonstrated notable improvements, boosting end-to-end resolution rates by up to 5.5% and reducing main-agent token consumption by up to 60% across SWE-bench Multilingual, SWE-bench Pro, and SWE-QA benchmarks, with minimal overhead.

Key takeaway

For Machine Learning Engineers developing or deploying LLM-powered coding agents, you should re-evaluate monolithic agent architectures. Integrating a dedicated, trained exploration subagent like FastContext can significantly reduce your main model's token consumption by up to 60% and improve task resolution rates by up to 5.5%. Consider modularizing your agent workflow to offload repository navigation to specialized, smaller models, thereby optimizing inference costs and enhancing overall performance.

Key insights

Dedicated exploration subagents improve LLM coding agent efficiency and accuracy by providing focused context.

Principles

• Repository exploration is a distinct, costly bottleneck for coding agents.
• Specialized models handle exploration more efficiently than general task solvers.
• Task-grounded RL effectively refines compact exploration models.

Method

FastContext delegates repository exploration, executing parallel Read, Glob, and Grep tool calls. It is trained via SFT for initial behavior and RL for task-relevant code localization.

In practice

• Implement a dedicated subagent for repository exploration to save main LLM tokens.
• Train specialized models (e.g., 4B-RL) for efficient code search.
• Utilize parallel tool calls for faster codebase analysis.

Topics

• FastContext
• Coding Agents
• Repository Exploration
• LLM Efficiency
• Reinforcement Learning
• SWE-bench

Code references

• microsoft/fastcontext
• BurntSushi/ripgrep
• THUDM/slime

Best for: AI Scientist, Machine Learning Engineer, AI Engineer

Related on AIssential

• SWE-Explore: Benchmarking How Coding Agents Explore Repositories — SWE-Explore benchmarks coding agents' line-level repository exploration, isolating it from end-to-end repair outcomes.
• Organize then Retrieve: Hierarchical Memory Navigation for Efficient Agents — HORMA uses a hierarchical, file-system-like memory to enable efficient, detailed context retrieval for LLM agents in long-horizon tasks.
• LLM with 12M Context Window — SubQ introduces subquadratic sparse attention, enabling massive context windows with significantly reduced compute and cost.
• Context Pruning for Coding Agents via Multi-Rubric Latent Reasoning — Decomposing code relevance into semantic and dependency rubrics improves LLM coding agent context pruning.
• Continually improving our agent harness — Optimizing LLM agent performance requires a highly customized, iteratively refined, and dynamically managed harness.
• LLM Agents Can See Code Repositories — Integrating visual structural context with text significantly boosts coding agent efficiency and accuracy in repository tasks.

Open in AIssential →

Editorial summary, takeaway, and curation by AIssential. Original article published by cs.SE updates on arXiv.org.