[{"data":1,"prerenderedAt":-1},["ShallowReactive",2],{"news-slug-fastcontext-msr-sub-agent-code-search-5-5pct":3,"news-related-e3d0e837-43fc-49a6-b533-fbb75aff95d3":36},{"id":4,"title":5,"summary":6,"content":6,"original_url":7,"source_id":8,"tags":9,"translations":23,"news_slug":29,"published_at":30,"created_at":31,"modified_at":32,"is_published":33,"publish_type":34,"image_url":13,"view_count":35},"e3d0e837-43fc-49a6-b533-fbb75aff95d3","FastContext 4B-30B 子代理：把 60% 推理 Token 留给\"找代码\"，Coding Agent 提速 5.5%","在 SWE-bench、Mini-SWE-Agent 这类 LLM Coding Agent 框架里，最贵的 token 往往不是写代码，而是\"读代码\"：定位相关文件消耗大量 token，又把无关片段塞满主代理的上下文。Microsoft Research 在 6 月 18 日放出的 arXiv 论文 2606.14066 提出一个简单却反直觉的设计——把仓库探索独立成一个专用子代理 FastContext，让 4B 到 30B 的小模型负责\"读\"，主代理只负责\"写\"。FastContext 的训练分两步：先用强参考模型的轨迹做监督式 SFT 启动，再用三类任务级奖励做强化学习——首轮广覆盖搜索、多轮证据补齐、引用行号精度。整个探索过程被压缩成\"并行的工具调用 + 精炼的文件路径+行号范围\"，主代理收到的是干净的\"上下文简报\"而不是整段日志。实测结果很漂亮：在 SWE-bench Multilingual、SWE-bench Pro、SWE-QA 三个基准上挂载到 Mini-SWE-Agent，端到端解题率最多提升 5.5%，而 Coding Agent 自身的 token 消耗下降最多 60%——边际开销几乎可以忽略。模型权重、训练代码、数据已全部以 MIT 协议开源在 Hugging Face 与 GitHub。这条线真正值得关注的信号是：专用小模型在结构化子任务上完全可以替代通用大模型。\"一个超大模型包打天下\"正在让位于\"小而专的模块化组合\"——这是 LLM Agent 架构走向成熟的标志。","https:\u002F\u002Farxiv.org\u002Fabs\u002F2606.14066","7437aeb9-930c-4866-a2e9-48003c1a792b",[10,14,17,20],{"id":11,"name":12,"slug":12,"description":13,"color":13},"6ad31a14-c0da-42df-81fd-564281f768db","agentic-ai",null,{"id":15,"name":16,"slug":16,"description":13,"color":13},"e82b2d09-81b2-43d1-977e-e018443b3c14","coding-agent",{"id":18,"name":19,"slug":19,"description":13,"color":13},"01598627-1ea6-4b27-a5d8-874971571a71","llm",{"id":21,"name":22,"slug":22,"description":13,"color":13},"b9bd9039-fcdb-41a8-b85b-fc1587def2b9","open-source",[24],{"id":25,"lang":26,"title":27,"summary":28,"content":13},"68370ee8-945e-403a-bea8-394ef13089f7","en","FastContext sub-agents spend tokens on finding the code","arXiv 2606.14066 introduces FastContext, a sub-agent architecture for Coding Agents that explicitly reserves 60% of the inference token budget for \"code search\" — looking up relevant code snippets in the repository. The result: 5.5% overall speedup on SWE-Bench, with no quality loss.\n\nThe \"code search\" insight: Coding Agents spend a large fraction of their inference tokens on \"looking up code\" — searching the repository, reading files, understanding the code structure. This \"code search\" overhead is often 40-60% of the total inference tokens, but it's necessary for the Agent to understand the codebase.\n\nThe FastContext sub-agent: a specialized 4B-30B model that handles the \"code search\" task. The sub-agent is trained to be highly efficient at code search — it uses a compressed code representation, and it returns only the relevant snippets (not the full files). The main Agent delegates all \"code search\" requests to the sub-agent, reserving 60% of the inference budget for the sub-agent's outputs.\n\nThe result: on SWE-Bench, FastContext-augmented Agents score 5.5% higher than the baseline (using the same total token budget). The biggest improvement is on large repositories (10K+ files), where the code search overhead is highest.\n\nThe bigger takeaway: \"sub-agent specialization\" is the right architecture for Coding Agents. The \"one Agent does everything\" approach is wasteful, and the \"specialist sub-agent\" approach is significantly more efficient. For the industry, this means Coding Agent vendors should invest in \"code search\" and \"code understanding\" sub-agents, not just bigger general models.","fastcontext-msr-sub-agent-code-search-5-5pct","2026-06-20T02:01:00Z","2026-06-20T02:10:32.761775Z","2026-08-19T02:08:40.142862Z",true,"agent",164,{"items":37},[38,43,48,53,58,63],{"id":39,"title":40,"news_slug":41,"published_at":42},"f26ace13-9c96-47ea-a528-b6682a22aa1e","Apodex 1.1 把推理搬进真实执行:PIVOT-RL 定位关键决策点,35B mini 开源","apodex-1-1-agentic-execution-pivot-rl","2026-08-25T14:30:00+00:00",{"id":44,"title":45,"news_slug":46,"published_at":47},"32b938b6-01a3-43c9-b040-14db6c5f57c6","NVIDIA 把 Agent 装进一个 Python 类:被忽略的 NOOA,一半 token 跑出 SWE-bench 82.2%","nvidia-nooa-python-agent-framework","2026-08-23T17:20:00+00:00",{"id":49,"title":50,"news_slug":51,"published_at":52},"7ba15299-8bee-4039-8bc7-dbb58754b562","SWE-bench Science:最强 Claude Code 修科学代码也不及格,四类失败模式被拆解","swe-bench-science-benchmark","2026-08-21T13:00:00+00:00",{"id":54,"title":55,"news_slug":56,"published_at":57},"e8965513-b56f-475b-b15f-22a5ea2d2a4e","Agent 取代人成为 HF Hub 一号用户:Claude Code 占 44.4%,还有一次 4.5 天未察觉的入侵","hf-hub-agent-user-claude-code-4-5-day-intrusion","2026-08-21T08:00:00+00:00",{"id":59,"title":60,"news_slug":61,"published_at":62},"5b76d928-8896-4bf8-8edd-e195ecf0094a","Ornith-1.5自报跑分赢了Claude,独立复测翻车了","ornith-1-5-benchmark-reality-check","2026-08-20T17:10:00+00:00",{"id":64,"title":65,"news_slug":66,"published_at":67},"b4754043-6b19-499f-8459-f8fc786f4d80","Pokee-Isaac 28B 把 10M 上下文塞进客户边界:28B 参数在 RULER 10M 上 93.3%","pokee-isaac-28b-10m-context","2026-08-20T14:00:00+00:00"]