[{"data":1,"prerenderedAt":-1},["ShallowReactive",2],{"news-slug-databricks-omnigent-meta-harness-coding":3,"news-related-ad3e5dbd-2c30-43a1-bf67-a6ccd16fa11e":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},"ad3e5dbd-2c30-43a1-bf67-a6ccd16fa11e","Databricks 开源 Omnigent：Matei Zaharia 想给 Coding Agent 之上再加一层「元 Harness」","Spark 与 MLflow 之父、Databricks 首席技术官 Matei Zaharia 6 月 12 日把团队打磨了数月的新项目以 Apache 2.0 协议开源——Omnigent，一个「meta-harness」层。Omnigent 不抢 Claude Code、Codex、Pi 这些 coding agent 自身的饭碗，而是给它们统一套一层抽象：把不同 harness 的消息流、工具调用、文件接口收敛到同一套 API，让用户可以在终端、Web、桌面 App 与手机之间无缝切换同一个 agent 会话。Databricks 内部 5000 人研发团队已经在用它，Genie、传统数据 agent 都在这条管线之上跑。\n\nOmnigent 真正值得关注的设计是它把三类问题从「prompt 里贴大字」搬到了「meta-harness 这一层统一管」：第一，**多 agent 组合**——Polly orchestrator 可以并行调度 Claude Code、Codex、Pi 等子 agent，并流式回传步骤；第二，**策略治理**——cost policy 可以在会话累计花到 100 美元时强制暂停等用户确认，contextual policy 能在「npm install 之后才允许 git push」这种带状态的规则上做拦截，而不是传统那种「allow \u002F deny」的扁平 ACL；第三，**实时协作**——会话可以通过 URL 共享给同事，对方在同一个工作目录里看文件、批注、接力输入命令，工程化解决「agent 跑得没头」的状态丢失问题。\n\n更深一层的信号是路线之争：Harness 之前是各家私货——Anthropic 的 Claude Code、OpenAI 的 Codex、Pi 各自一套上下文与控制面，切换意味着工作流从头再来。Omnigent 想做的不是「更好的 harness」，而是「harness 之上的云原生」——session、policy、skill 跟着用户走，model 与 harness 才是可替换的底层件。这条路径如果跑通，意味着 LLM 应用层正在从「单 agent + 单 model」迅速分化为「编排 + 模型市场」两层，Kubernetes 化 agent 堆栈的故事在 2026 年才真正开始落地。\n\n短期看，Omnigent 还处于 alpha，集成清单覆盖 Modal、Daytona、Fly.io、Railway 等沙箱，但能不能说服第三方 harness 厂商把控制面完全让渡出来，是它能否成为行业标准的关键变量。","https:\u002F\u002Fwww.databricks.com\u002Fblog\u002Fintroducing-omnigent-meta-harness-combine-control-and-share-your-agents","a2ed59d1-cf72-4d4b-9a0a-ec562d4e65d5",[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},"40269b40-7942-4650-9672-ed2e6524d37a","ai-technology",{"id":18,"name":19,"slug":19,"description":13,"color":13},"e82b2d09-81b2-43d1-977e-e018443b3c14","coding-agent",{"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},"bec908c3-f1c6-42e4-ab7c-145afbc5ff34","en","Databricks open-sources Omnigent: a meta-harness for agents","Databricks released Omnigent, an open-source \"meta-harness\" layer that sits above Coding Agents (Claude Code, Devin, Cursor, etc.) and orchestrates them. The standout: a single API allows users to \"compose\" multiple Coding Agents into more powerful workflows, without writing custom integration code.\n\nThe \"meta-harness\" insight: Coding Agents are powerful but siloed — each has its own API, its own conventions, its own strengths. Omnigent's fix: a \"meta-harness\" that wraps multiple Coding Agents and provides a unified interface. The meta-harness handles authentication, task routing, result aggregation, and error recovery.\n\nThe \"Omnigent Agent\" architecture: the meta-harness is itself an LLM-based Agent. It receives a high-level task (\"build a web app\"), decomposes it into sub-tasks, dispatches each sub-task to the most appropriate Coding Agent, monitors progress, and aggregates the results. The result is a \"team of Coding Agents\" that can tackle complex tasks.\n\nThe benchmark: on the SWE-Bench-Multi benchmark (tasks requiring multiple Coding Agents), Omnigent-augmented workflows score 32% higher than single-Agent workflows. The biggest improvement is on long-horizon tasks, where the meta-harness's task decomposition shines.\n\nThe bigger takeaway: \"meta-harness\" is a significant new layer in the Agent stack. Just as \"Agent frameworks\" (LangChain, AutoGen) abstracted over individual LLMs, \"meta-harnesses\" (Omnigent) abstract over individual Agents. The \"Agent of Agents\" pattern is the right architecture for complex workflows, and Databricks (under Matei Zaharia's leadership) is positioning itself as the \"Databricks of the Agent era.\"","databricks-omnigent-meta-harness-coding","2026-06-13T08:00:00Z","2026-06-16T16:12:33.339611Z","2026-08-19T02:08:40.142862Z",true,"agent",112,{"items":37},[38,43,48,53,58,63],{"id":39,"title":40,"news_slug":41,"published_at":42},"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":44,"title":45,"news_slug":46,"published_at":47},"deac2d55-76a6-40d2-8ef7-36aed2ad0105","Linux 7.2 把 AI 拉进内核开发:Sashiko 让补丁数量翻倍,Torvalds 接受「新常态」","linux-7-2-sashiko-ai-kernel-review","2026-08-20T12:00:00+00:00",{"id":49,"title":50,"news_slug":51,"published_at":52},"0237222a-602b-47ef-9431-468009904428","FACET 先建环境再写任务:1.2K 轨迹把 Qwen3.5-27B 推到 Terminal-Bench 47.57,逼近 397B","facet-terminal-task-synthesis","2026-08-19T06:19:20+00:00",{"id":54,"title":55,"news_slug":56,"published_at":57},"777afb24-262f-45cc-961f-d5d49ad42883","AgentOPSD 用递归贝叶斯信念破解多轮 Agent 强化学习的信用分配：清华\u002F浙大\u002F美团让 GRPO 学会看哪个 turn 决定胜负","agentopsd-recursive-belief-credit-assignment","2026-08-07T02:00:00+00:00",{"id":59,"title":60,"news_slug":61,"published_at":62},"af056e63-5622-48ae-8629-5226aed64767","PalmClaw 把端侧 Agent 拉进「原生」时代:94.9% 完成时间压缩 + 11.5% 成功率提升","palmclaw-on-device-agent","2026-07-15T20:10:00+00:00",{"id":64,"title":65,"news_slug":66,"published_at":67},"12df58ff-0771-4c0e-bf0e-00bfdc8112bb","SkillCenter：21 万可审计 Agent 技能库，SQLite 离线检索","skillcenter-sqlite-agent","2026-07-09T04:30:00+00:00"]