FastContext

Sources#

• FastContext: Training Efficient Repository Explorer for Coding Agents

Summary#

FastContext is an open-source repository-exploration subagent for coding agents, from Shaoqiu Zhang, Maoquan Wang, Yuling Shi et al. at Microsoft CoreAI and Shanghai Jiao Tong University (arXiv 2606.14066, June 2026; corresponding author Shengyu Fu). It is the system instance of the Repository Exploration Subagent thesis: a dedicated, read-only helper the main agent invokes on demand to locate code, returning compact path:line-range citations instead of a long exploratory trail. Integrated into Mini-SWE-Agent across GPT-5.4, GLM-5.1, and Kimi-K2.6 main agents, it improves end-to-end resolution up to 5.5% while cutting main-agent token consumption up to 60%. Code and data: github.com/microsoft/fastcontext.

What it is#

A runtime CLI helper (fastcontext -q "..." --format concise) that runs in the same task container as the main agent. It exposes exactly three language-agnostic read-only tools — READ, GLOB, GREP (ripgrep) — issues parallel tool calls per turn, and emits a single <final_answer> block of file paths + line ranges. It cannot edit files or submit patches; only its final evidence block enters the main-agent context, while its intermediate observations go to separate logs.

Model family#

Specialized explorers trained on Qwen backbones, spanning 4B–30B:

Variant	Backbone	Training	Role
FC-4B-SFT	Qwen3-4B-Instruct	SFT only	compact baseline
FC-4B-RL	Qwen3-4B-Instruct	SFT → GRPO RL	deployment target — beats 30B-SFT on several settings
FC-30B-SFT	Qwen3-Coder-30BA3B	SFT only	scaling reference

The 4B models are the deployment targets because exploration must be cheap enough to run routinely; 4B-RL is the headline result — task-grounded RL makes a compact explorer competitive without an expensive 30B RL run.

Training stack#

• SFT data: 2,954 filtered traces from Sonnet 4.6 (Anthropic) as the reference model, split into parallel_toolcalls / multiturn_traj / linerange sources. Trained on the Slime/Megatron stack, 3 epochs, assistant-token-only loss, 128K context.
• RL: GRPO over 400 prompts / 395 repos, reward = patch-derived file/line F1 + bounded-parallelism bonus + format penalty; rollouts via SGLang (thinking disabled, up to 8 turns, 16 trajectories/prompt).

Significance#

FastContext is the open, published counterpart to the proprietary subagent mechanisms inside Claude Code, Codex, GitHub Copilot CLI, and Cursor — its stated motivation is that those exploration mechanisms are closed, leaving the research community without an open training/evaluation recipe. Its sharpest argument is that the architectural separation of exploration from solving carries much of the benefit even without a trained model (the "same-model exploration" baseline already helps), making exploration a modular, optimizable, separately-evaluable stage of a coding agent.

Connections#

• Repository Exploration Subagent — the concept FastContext instantiates (bottleneck analysis, delegation contract, training recipe, results)
• Claude Code — the proprietary subagent analog FastContext open-sources
• Anthropic — Sonnet 4.6 is the reference model whose traces seed FastContext's SFT corpus
• Symphony — a sibling modular agent system (ticket-driven orchestration vs. exploration delegation)
• Client-Side Agent Optimization — FC's train-a-bespoke-small-model-per-role move extends AgentOpt's model-per-role optimization

Open Questions#

• Can the SFT+RL recipe push the explorer below 4B (1.7B / 0.6B) and make exploration effectively free?
• Does the gain transfer beyond Mini-SWE-Agent to richer harnesses with their own subagent orchestration?

Sources#

• FastContext: Training Efficient Repository Explorer for Coding Agents