FastContext: Give the Coding Agent a Scout
Microsoft and Shanghai Jiao Tong have a clean idea for making coding agents both better and cheaper: stop forcing the main agent to read the entire repo. FastContext, which hit 312 upvotes on HuggingFace with code at github.com/microsoft/fastcontext, trains a dedicated exploration subagent whose only job is to go find the relevant code, fire off searches in parallel, and hand back tight file-and-line citations, instead of dumping a pile of exploratory reads into the main agent's context window.
The numbers are the whole argument. Bolted onto Mini-SWE-Agent, it lifts end-to-end resolution by up to 5.5 points while cutting the main agent's token consumption by up to 60 percent, measured across SWE-bench Multilingual, SWE-bench Pro, and SWE-QA. They trained the explorer at sizes from 4B to 30B using supervised fine-tuning plus reinforcement learning. Think of it like sending a scout ahead instead of marching the whole army through every alley. The general only ever sees the part of the map that matters.
This is the lesson the field keeps relearning from different angles: context is the scarce resource, and the agent that wastes the least of it wins. We've watched this theme run through turbovec, through the repo-exploration work, through Is Grep All You Need. FastContext is the cleanest separate-search-from-solving datapoint yet, and the reason it matters is that it ships as a trainable subagent, not a prompt trick. The skill lives in weights you can reuse, not in a clever instruction you paste every time. Paper: https://huggingface.co/papers/2606.14066
← Back to all articles
The numbers are the whole argument. Bolted onto Mini-SWE-Agent, it lifts end-to-end resolution by up to 5.5 points while cutting the main agent's token consumption by up to 60 percent, measured across SWE-bench Multilingual, SWE-bench Pro, and SWE-QA. They trained the explorer at sizes from 4B to 30B using supervised fine-tuning plus reinforcement learning. Think of it like sending a scout ahead instead of marching the whole army through every alley. The general only ever sees the part of the map that matters.
This is the lesson the field keeps relearning from different angles: context is the scarce resource, and the agent that wastes the least of it wins. We've watched this theme run through turbovec, through the repo-exploration work, through Is Grep All You Need. FastContext is the cleanest separate-search-from-solving datapoint yet, and the reason it matters is that it ships as a trainable subagent, not a prompt trick. The skill lives in weights you can reuse, not in a clever instruction you paste every time. Paper: https://huggingface.co/papers/2606.14066
Comments