每日 Harness 开源 · Source
全部刊期 · All issues

每日 Harness

2026-08-08 · Saturday, August 8, 2026

智能体走向可测可控高效

视图 · View

今日重点 · Today's Highlights

OrchestraBench1 - 用可复现故障注入把多智能体编排拆成路由、传播与恢复问题，测出的级联半径会随工作流深度从 0.9 扩到 4.7。

全文 ↓

CodeGrep2 - 14B 专职检索 agent 在 SWE-Bench Verified 全集同时提高解题率并削减已解决任务 19% 的 token 消耗。

全文 ↓

Oans3 - 通过跳过已知区间与流水预取重做文件去重热路径，207 万文件的热重跑由约 11 分钟缩至 92 秒。

全文 ↓

DwarfStar / ds45 - 为路由 MoE 模型从加载器到 coding agent 重做整条本地推理栈，8×L40S 自测达到聚合 120 tok/s 生成吞吐。

全文 ↓

论文 · Papers

15 项 · 论文

开源 / 项目 · Projects

15 项 · 开源 / 项目

Alyph19alyph.ai原文 ↗

上下文工程框架与脚手架其他垂直

Alyph 把 LLM 对话历史呈现为可编辑、可分叉的上下文图，用户能从任意节点产生并行分支，再比较不同提示或模型响应，而不必复制整段会话。

–

Coarena21coarena.ai原文 ↗

基准计算机·Web

Coarena 让社区提交真实计算机操作任务，再在相同任务上比较模型执行表现，借此让评测集随用户场景持续生长。

–

DayOtter23github.com原文 ↗

github.com

DayOtter 统一 Google、Microsoft 365、CalDAV 与 ICS 日历，提供预约、专注时段、冲突处理、团队能力和 Stripe；自然语言、语音、WhatsApp、SMS 助手只提出变更，确认后才落盘。

–

Bifrost25github.com原文 ↗

github.com

Bifrost 以 OpenAI 兼容 API 统一 23 家以上模型供应商，并在同一服务中实现负载均衡、故障转移、语义缓存、插件治理与 MCP 网关。

–

Line9 Diagram26line9.ai原文 ↗

line9.ai

Line9 Diagram 保留 Mermaid flowchart 的文本输入，替换自动布局与渲染层，目标是减少大型流程图中的交叉、拥挤和手工位置修正。

–

local_bench27github.com原文 ↗

github.com

local_bench 用单命令 Python TUI 自动发现 Ollama、LM Studio、llama.cpp、vLLM 和 OpenAI 兼容服务，测量 tok/s、首 token 延迟、内存与可选 joules/token。

–

本期重点Oans3github.com原文 ↗

github.com

Oans 在 duperemove 基础上先跳过已散列或已经共享的区间，再用流式管线与页缓存预取降低无效 I/O，最终调用 FIDEDUPERANGE，由内核逐字节确认后共享物理块。

–

Llmem29github.com原文 ↗

Agent 记忆检索与知识接地编码

Llmem 以约一万行 Go 和单个 SQLite 文件实现 coding-agent 记忆，不依赖向量数据库、嵌入 API 或云服务；检索采用 BM25，可选 WordNet 扩词。

–

ARF31korvo.xyz原文 ↗

korvo.xyz

ARF 试图用文件格式记录 AI evaluation run，把输入、配置、模型输出、评分和运行环境打包为可交换工件，并从同一记录生成可复现摘要。

–

本期重点Pokémon Emerald on RP23504github.com原文 ↗

github.com

项目把反编译的 ARMv4T 游戏逻辑重编译到 252MHz Cortex-M33，核心 0 运行游戏，核心 1 用软件模拟 GBA PPU，11.7MB 游戏从 16MB QSPI 直接执行。

–

行业动态 · Industry News

15 项 · 行业动态

DeepSeek V4 Flash 073132arcprize.org原文 ↗

arcprize.org

ARC Prize 发布 DeepSeek V4 Flash 0731 在 ARC-AGI 抽象网格任务上的独立结果，把新模型放入统一的非语言推理协议比较。

–

博客文章 · Blog Posts

15 项 · 博客文章

Nvidia’s Vera Whitepaper Has a Thread Loose50chipsandcheese.com原文 ↗

chipsandcheese.com

文章逐项核对 Nvidia Vera 白皮书的处理器结构与性能表述，把缓存、内存、核心组织和比较基线重新放进共同约束，指出若干数字或因果解释无法顺接。

–

Taste Is All That's Left51notashelf.dev原文 ↗

notashelf.dev

文章认为生成候选实现的成本快速下降后，稀缺能力转向定义问题、筛选结果、决定何时停止，以及主动舍弃不合适的可能性。

–

How a device finds encrypted DNS by itself55blog.dundns.eu原文 ↗

blog.dundns.eu

文章解释 Discovery of Designated Resolvers：设备从网络已有解析器身份出发，发现其声明的 DoH/DoT 端点，并验证明文服务与加密端点的升级关系。

–

A new model for source control56wiki.alcidesfonseca.com原文 ↗

wiki.alcidesfonseca.com

文章尝试重新组织版本、变更与协作关系，避免传统分支、提交和合并把独立意图压入一条难拆解的历史。

–

Nix Evaluation Is a Scheduling Problem57notashelf.dev原文 ↗

notashelf.dev

文章把 Nix 求值视为依赖调度：表达式可能阻塞、展开或复用，运行时必须选择任务粒度和并行顺序，而非只提升解释器某个函数的速度。

–

Möbius-Strip Crosswords60quuxplusone.github.io原文 ↗

quuxplusone.github.io

文章设计首尾按 Möbius 带反向粘合的填字网格，使跨边界单词的方向与邻接遵循一次扭转后的拓扑规则。

–

Sidebar: building a private ESP32 intercom61sundaradnus.ca原文 ↗

sundaradnus.ca

作者用 5 台 ESP32-S3 触屏设备搭建儿童对讲系统，不引入账户、信息流或云服务，把设备发现、语音和家庭网络交互限制在自有硬件内。

–

引用来源 · References

72 条 · 引用
  1. 1 OrchestraBench: Evaluating Multi-Agent Orchestration Failure Modes, Recovery, and Decomposition Quality. arXiv:2608.05263https://arxiv.org/abs/2608.05263 ↩ 回到正文 · back to text
  2. 2 CodeGrep: An RL-Trained Retrieval Agent for LLM Coding Agents. arXiv:2608.05886https://arxiv.org/abs/2608.05886 ↩ 回到正文 · back to text
  3. 3 Oans. GitHub: martinus/oanshttps://github.com/martinus/oans ↩ 回到正文 · back to text
  4. 4 Pokémon Emerald on RP2350. GitHub: mattdeeds/pokeemerald-rp2350https://github.com/mattdeeds/pokeemerald-rp2350 ↩ 回到正文 · back to text
  5. 5 DwarfStar / ds4. GitHub: antirez/ds4https://github.com/antirez/ds4 ↩ 回到正文 · back to text
  6. 6 SkillTrace: Multi-Trace Provenance Auditing for LLM-Agent Skill Reuse. arXiv:2608.05204https://arxiv.org/abs/2608.05204 ↩ 回到正文 · back to text
  7. 7 SearchAuditor: Auditing and Attributing Failures in Long-Horizon Search Agents. arXiv:2608.05212https://arxiv.org/abs/2608.05212 ↩ 回到正文 · back to text
  8. 8 EcoAgent-Bench: Evaluating Economic Decision-Making in Budget-Constrained LLM Agents. arXiv:2608.05519https://arxiv.org/abs/2608.05519 ↩ 回到正文 · back to text
  9. 9 SkillTV-Bench: Benchmarking How Well Judges Perform on Skill-Augmented Agentic Execution. arXiv:2608.05573https://arxiv.org/abs/2608.05573 ↩ 回到正文 · back to text
  10. 10 SkillHEX: Improving Agent Skills via Hypothesis-Driven Autonomous Exploration and Exploitation. arXiv:2608.05628https://arxiv.org/abs/2608.05628 ↩ 回到正文 · back to text
  11. 11 DreamGuard: Efficient Runtime Guardrail for LLM Agents via Risk-Aware World Model. arXiv:2608.05695https://arxiv.org/abs/2608.05695 ↩ 回到正文 · back to text
  12. 12 When Self-Evolution Backfires: Pre-Commit Gating against Skill Contamination in LLM Agents. arXiv:2608.05810https://arxiv.org/abs/2608.05810 ↩ 回到正文 · back to text
  13. 13 Contextual Information Policy Optimization for Search Agents. arXiv:2608.06128https://arxiv.org/abs/2608.06128 ↩ 回到正文 · back to text
  14. 14 Comparative Approaches to Agent Retrieval over Large Skill Libraries. arXiv:2608.06196https://arxiv.org/abs/2608.06196 ↩ 回到正文 · back to text
  15. 15 HarnessOpt-Bench: Evaluating LLMs at Harness Optimization. arXiv:2608.06301https://arxiv.org/abs/2608.06301 ↩ 回到正文 · back to text
  16. 16 TRAJDEBUG: Tracing Error Lifecycle to Identify Critical Failures in Long-Horizon Agent Trajectories. arXiv:2608.06346https://arxiv.org/abs/2608.06346 ↩ 回到正文 · back to text
  17. 17 SkillZip: Contract-Preserving Graph Compression for Scalable Agent Skill Libraries. arXiv:2608.05604https://arxiv.org/abs/2608.05604 ↩ 回到正文 · back to text
  18. 18 Screenshots or Tools? Eliciting Tool Use and Managing Multimodal Context in Hybrid GUI-MCP Computer-Use Agents. arXiv:2608.03327https://arxiv.org/abs/2608.03327 ↩ 回到正文 · back to text
  19. 19 Alyphhttps://www.alyph.ai/ ↩ 回到正文 · back to text
  20. 20 Zaivern Code. GitHub: tacyan/zaivern-codehttps://github.com/tacyan/zaivern-code ↩ 回到正文 · back to text
  21. 21 Coarenahttps://coarena.ai ↩ 回到正文 · back to text
  22. 22 hpc-sandbox-benchmarks. GitHub: starslingdev/hpc-sandbox-benchmarkshttps://github.com/starslingdev/hpc-sandbox-benchmarks ↩ 回到正文 · back to text
  23. 23 DayOtter. GitHub: Dayotter/dayotterhttps://github.com/Dayotter/dayotter ↩ 回到正文 · back to text
  24. 24 MCP Debug Proxy. GitHub: ytkoka/mcp-debug-proxyhttps://github.com/ytkoka/mcp-debug-proxy ↩ 回到正文 · back to text
  25. 25 Bifrost. GitHub: maximhq/bifrosthttps://github.com/maximhq/bifrost/ ↩ 回到正文 · back to text
  26. 26 Line9 Diagramhttps://line9.ai/diagram ↩ 回到正文 · back to text
  27. 27 local_bench. GitHub: SupermodularAI/local_benchhttps://github.com/SupermodularAI/local_bench ↩ 回到正文 · back to text
  28. 28 Mirafoldhttps://mirafold.com/ ↩ 回到正文 · back to text
  29. 29 Llmem. GitHub: netrixone/llmemhttps://github.com/netrixone/llmem ↩ 回到正文 · back to text
  30. 30 Remembrane. GitHub: satyasairay/remembranehttps://github.com/satyasairay/remembrane ↩ 回到正文 · back to text
  31. 31 ARFhttps://www.korvo.xyz/arf ↩ 回到正文 · back to text
  32. 32 DeepSeek V4 Flash 0731https://arcprize.org/results/deepseek-v4-flash-0731 ↩ 回到正文 · back to text
  33. 33 Oracle bans AI-generated code from OpenJDKhttps://app.dealroom.co/news/feed/oracle-bans-ai-generated-code-from-openjdk-despite-ellison-s-claim-oracle-isn-t-writing-its-own-code ↩ 回到正文 · back to text
  34. 34 Databricks drove down AI coding spend 70%https://www.databricks.com/blog/managing-ai-coding-costs-scale ↩ 回到正文 · back to text
  35. 35 Responding to the next frontier of critical cyber capabilitieshttps://openai.com/index/responding-next-frontier-critical-cyber-capabilities/ ↩ 回到正文 · back to text
  36. 36 New Mexico court orders Meta to pay $567m over harms to children’s mental healthhttps://www.theguardian.com/technology/2026/aug/06/new-mexico-court-meta ↩ 回到正文 · back to text
  37. 37 AMD acquires Taalas to boost inference performance by etching models in siliconhttps://www.theregister.com/systems/2026/08/06/amd-acquires-ai-chip-startup-taalas-to-boost-inference-performance-by-etching-models-into-silicon/5284344 ↩ 回到正文 · back to text
  38. 38 GitHub Actions and Pages are experiencing degraded availabilityhttps://www.githubstatus.com/incidents/qcvjkzcs7j74 ↩ 回到正文 · back to text
  39. 39 Improving GPT‑5.6 Sol in ChatGPT, expanding GPT‑5.6 Luna access for free usershttps://openai.com/index/improving-gpt-5-6-sol-in-chatgpt/ ↩ 回到正文 · back to text
  40. 40 Kitesurf: Agent-first browser that runs in V8 isolateshttps://blog.cloudflare.com/kitesurf/ ↩ 回到正文 · back to text
  41. 41 Framework discloses data breach via Metabase 0-dayhttps://community.frame.work/t/framework-data-breach-discussion/83939 ↩ 回到正文 · back to text
  42. 42 UAD issue #1426: feat(pkg-state): disable by defaulthttps://github.com/Universal-Debloater-Alliance/universal-android-debloater-next-generation/issues/1426 ↩ 回到正文 · back to text
  43. 43 2027 memory capacity is reportedly sold outhttps://www.ign.com/articles/ramageddon-continues-another-year-as-2027-memory-capacity-is-reportedly-sold-out ↩ 回到正文 · back to text
  44. 44 USA Today Co. partners with Palantir to analyze audience datahttps://www.niemanlab.org/2026/08/americas-largest-newspaper-chain-usa-today-co-partners-with-palantir-to-analyze-audience-data/ ↩ 回到正文 · back to text
  45. 45 New Orleans is testing Carbyne’s AI-powered Emergency Call Triage softwarehttps://www.shreveporttimes.com/story/news/local/louisiana/2026/07/28/is-new-orleans-using-ai-to-answer-911-calls-instead-of-human-dispatchers-impacts-emergencies-crime/91065014007/ ↩ 回到正文 · back to text
  46. 46 App Store Rejection of the Week: Dark Hourshttps://daringfireball.net/2026/08/app_store_rejection_of_the_week_dark_hours ↩ 回到正文 · back to text
  47. 47 Making Postgres 300x faster for analytics: batching, operator fusion, and SIMDhttps://malisper.me/how-we-made-postgres-hundreds-of-times-faster-the-query-engine/ ↩ 回到正文 · back to text
  48. 48 What happens if an entire class of workers loses faith in their careershttps://www.noemamag.com/why-is-everyone-in-tech-so-sad/ ↩ 回到正文 · back to text
  49. 49 A year of fighting scrapers on my 1.5 million-page websitehttps://patronview.com/news/99-percent-of-my-website-traffic-is-bots/ ↩ 回到正文 · back to text
  50. 50 Nvidia’s Vera Whitepaper Has a Thread Loosehttps://chipsandcheese.com/p/nvidias-vera-whitepaper-has-a-thread ↩ 回到正文 · back to text
  51. 51 Taste Is All That's Lefthttps://notashelf.dev/posts/taste-is-all-thats-left ↩ 回到正文 · back to text
  52. 52 Software development with AI is starting to feel like cooking steakhttps://blog.sydorets.com/en/posts/almost-no-skill-required-to-cook-a-steak/ ↩ 回到正文 · back to text
  53. 53 Humans missed 1 in 3 threats approving AI agent commands across 40k game runshttps://scalex.dev/blog/ai-agent-permissions-stats/ ↩ 回到正文 · back to text
  54. 54 AI-generated vulnerability patches require human reviewhttps://1password.com/blog/why-ai-generated-patches-still-require-human-review ↩ 回到正文 · back to text
  55. 55 How a device finds encrypted DNS by itselfhttps://blog.dundns.eu/posts/ddr-encrypted-dns-discovery/ ↩ 回到正文 · back to text
  56. 56 A new model for source controlhttps://wiki.alcidesfonseca.com/blog/a-new-model-for-source-control/ ↩ 回到正文 · back to text
  57. 57 Nix Evaluation Is a Scheduling Problemhttps://notashelf.dev/posts/evix ↩ 回到正文 · back to text
  58. 58 The web server deployment model breaks at hobby scalehttps://w.on-t.work/web-deployment-model ↩ 回到正文 · back to text
  59. 59 From constraint models to playable puzzle gameshttps://zayenz.se/blog/post/constraint-generated-puzzle-games/ ↩ 回到正文 · back to text
  60. 60 Möbius-Strip Crosswordshttps://quuxplusone.github.io/blog/2026/08/04/mobius-crossword/ ↩ 回到正文 · back to text
  61. 61 Sidebar: building a private ESP32 intercomhttps://www.sundaradnus.ca/writing/sidebar-building-a-ham-radio-for-my-son ↩ 回到正文 · back to text
  62. 62 authentik. GitHub: goauthentik/authentikhttps://github.com/goauthentik/authentik ↩ 回到正文 · back to text
  63. 63 Frappe HRMS. GitHub: frappe/hrmshttps://github.com/frappe/hrms ↩ 回到正文 · back to text
  64. 64 tuicr. GitHub: agavra/tuicrhttps://github.com/agavra/tuicr ↩ 回到正文 · back to text
  65. 65 faceswap. GitHub: deepfakes/faceswaphttps://github.com/deepfakes/faceswap ↩ 回到正文 · back to text
  66. 66 Embabel Agent. GitHub: embabel/embabel-agenthttps://github.com/embabel/embabel-agent ↩ 回到正文 · back to text
  67. 67 Warpgate. GitHub: warp-tech/warpgatehttps://github.com/warp-tech/warpgate ↩ 回到正文 · back to text
  68. 68 Zola. GitHub: getzola/zolahttps://github.com/getzola/zola ↩ 回到正文 · back to text
  69. 69 Rayhunter. GitHub: EFForg/rayhunterhttps://github.com/EFForg/rayhunter ↩ 回到正文 · back to text
  70. 70 Guava. GitHub: google/guavahttps://github.com/google/guava ↩ 回到正文 · back to text
  71. 71 Nub. GitHub: nubjs/nubhttps://github.com/nubjs/nub ↩ 回到正文 · back to text
  72. 72 AutoGPT. GitHub: Significant-Gravitas/AutoGPThttps://github.com/Significant-Gravitas/AutoGPT ↩ 回到正文 · back to text