Most coding agents mix “search the repo” with “solve the task,” which wastes tokens and pollutes the solver's context. FastContext flips that convention: it is a dedicated, read-only exploration subagent that the main coding agent calls when it needs focused repository evidence. The core insight is simple but practical — delegate broad discovery to a compact explorer so the solver consumes fewer tokens and operates on succinct, high‑precision citations.
Key Capabilities
- Parallel read-only tools: issues multiple READ / GLOB / GREP calls in the same exploration turn to cover complementary hypotheses and reduce round trips.
- Compact, actionable evidence: returns file paths with line ranges (not long snippets), letting the main agent reason and edit with minimal context overhead.
- Training recipe tuned for exploration: bootstrapped with supervised fine-tuning on explorer trajectories and refined with task-grounded RL to optimize file/line-level citation F1 and bounded parallelism.
- Lightweight deployments and scale points: available as a 4B explorer (SFT/RL variants) and larger references; supports long contexts (up to 262K tokens) to handle big repositories.
Who It's For and Trade-offs
Great fit if you build or integrate LLM-based coding agents and want to offload repository navigation: reduces token consumption, improves evidence precision, and can raise end-to-end resolution rates when wired into a solver. Look elsewhere if your workflow requires the explorer to modify files (FastContext is read-only), if you need a single model that both explores and solves without extra orchestration, or if you cannot adopt the required subagent contract and tooling (READ/GLOB/GREP) in your runtime. Integration requires the main agent to accept compact file-line citations and perform subsequent reads/edits itself.
