FastContext: A specialized exploration agent reduces coding agent token usage by 60%

A major bottleneck for LLM coding agents is repository exploration: locating relevant code consumes substantial token budget and pollutes the agent’s context with irrelevant snippets. Traditional agents use the same model for both exploration and solving, leaving exploratory reads in the solver’s history. FastContext introduces a specialized exploration subagent that separates these tasks. Invoked on demand, it issues parallel tool calls and returns concise file paths and line ranges, providing focused context for the solving agent. FastContext is powered by models ranging from 4B to 30B parameters, bootstrapped from strong reference-model trajectories and refined with rewards for broad first-turn search, multi-turn evidence gathering, and precise citation generation.

Across SWE-bench Multilingual, SWE-bench Pro, and SWE-QA benchmarks, integrating FastContext into Mini-SWE-Agent improves end-to-end resolution rates by up to 5.5% while reducing overall coding-agent token consumption by up to 60%, with only marginal overhead. These results demonstrate that repository exploration can be effectively delegated to specialized models, significantly improving efficiency and performance for software engineering tasks.