In a nutshell: Dedicated exploration models (4B–30B parameters) can handle code search in repositories more efficiently than general solver models while significantly reducing context pollution.
Researchers from Microsoft present FastContext, a specialized subagent for LLM-based coding agents that separates codebase exploration from actual problem-solving and thereby reduces token consumption by up to 60 percent.
The problem is well known: when LLM-based coding agents are tasked with solving a software engineering project, they waste considerable token budgets on repository exploration. Irrelevant code snippets end up in the context history of the same model that later implements the solution — an architectural weakness that impairs both efficiency and solution quality.
FastContext separates these tasks: a dedicated, specialized exploration subagent handles codebase navigation, while the solver model works uncluttered on the solution. The exploration agent operates with parallel calls and returns precise file paths and line ranges — not complete snippets, but focused references. The system is based on trained models ranging from 4 billion to 30 billion parameters, bootstrapped from strong reference trajectories and refined via task-specific rewards for three core capabilities: broad first-turn search, multi-step evidence gathering, and precise citation generation.
Evaluated on SWE-bench Multilingual, SWE-bench Pro, and SWE-QA, FastContext integrated into Mini-SWE-Agent achieves improvements in end-to-end success rates of up to 5.5 percent while simultaneously reducing coding agent token consumption by up to 60 percent. The overhead for FastContext remains marginal. These results demonstrate that repository exploration can be treated as a separate, specialized task.
Code and training data are available on GitHub under microsoft/fastcontext.
Source: arxiv.org · Published June 11, 2026
Lumi AI News — AI-assisted curation pursuant to Art. 50 EU AI Act. Paraphrase and classification by Lumi News Pipeline v1.7.1.