AI 模型和平台
线性 B 发布代理工作室，提供编码 Agents 工程背景
在 Google 上的首选来源中添加 Unite.AI
LinearB拥有发布 Agentic Studio，将软件交付数据转化为工程领导和编码agents可以用来做决定的背景。 宣布 November 2，该平台将自然语言分析与跨越工具团队用于规划，写作，审查和发布软件的工程上下文层相结合。
发布反映了AI辅助开发中一个日益严重的问题：生产更多的代码并不能自动帮助一个组织更快地交付有用的软件。 审查可能变得拥挤，变化会变得难以评估，agents可以开始一项任务，而不知道哪些服务是脆弱的，或者为什么某个特定的固定事项。
根据公告，代理工作室向所有LinearB 客户开放。 Agent技能，包括将agent的努力与任务相匹配的初始能力，已经提前进入。
将传送数据输入工程背景
LinearB收集跨软件发送的信息，并将其正常化为共同模型。 称公司调和了Git和项目管理系统的工程师身份，将票面的变化连接到发行，新增 AI归属和基准测试帮助领导者解释结果。
重要步骤是连接这些记录。 牵引请求可以孤立地看成是直截了当的，同时会影响最近发生的事故、频繁的重新工作或复杂的所有权。 将变化与历史联系起来，会给一个审查者或编码agent信息，而仅凭一个寄存器快照可能无法提供这些信息。
在新的工作室中，领导者可以用简单的语言提问，获得建议的下一步，并利用AI的协助来创建度量衡，仪表板和其他文物。 LinearB 称组织可以定义平台内包含角色访问功能的用户可用的上下文。
这一方向符合公司的环境工程指导，它强调选择相关信息，而不是不加区别地填补agent的背景窗口。 它的框架考虑代码，改变背后的意图，跨系统的连接以及agent可以访问的边界。
匹配风险的早期获取技能
第一种agent技能旨在决定编码任务需要付出多少努力。 线性B 称考虑业务关联性，复杂性和风险，使用复工率，近期服务事件和所有权历史等信号。
低风险的改变可以得到较少的代币的较轻的处理，而更相应的改变则可以得到更深层次的分析和测试。 工程领导者可以看到任务被引导到低，中或高强度工作的频率。 公司称这一技能与任何编码的agent合作，尽管公告并没有提供特定agents的比较结果或量化实现的象征性节省。
即使在确定费用索赔之前，工程设想也是有用的。 统一的推理预算处理轻微的维修变化和敏感的生产固定。 针对具体任务的背景有助于在错误代价高昂时给予更多的关注，同时减少关于更简单变革的不必要的工作。
MCP 连接数据到现有的 AI 工作流程
线性B也通过模型背景协议(Model Context Protocol)，或MCP，曝光工程信息。 其MCP产品页面介绍了自然语言获取交付量度表，定制报告和瓶颈分析，以及审查自动化和路由等建议。
这让团队可以调查来自AI客户端的工程数据，而不是在仪表板之间手动移动。 MCP提供了与工具和信息的连接；上下文层提供了使关于寄存器或团队的问题有意义的具体交付关系。
作为新平台推出，解决是有区别的。 虽然发布描述了基于角色的接入和企业护栏，LinearB的公用MCP FAQ仍然将权限范围MCP接入列为路线图项目。 因此，新工作室的访问控制声明不应被解读为确认每个现有的MCP客户端已经继承了每一个LinearB权限。
为什么更多的AI输出罐会展出下游瓶形
LinearB的公告引用了2026 AI 基准测试从253个组织的270万拉力请求和83000名开发者中提取的。 据该公司称，至少75%的编码日使用AI的开发者将每周比一年前多96%的拉力请求合并。
这些数字描述了观察到的活动，而不是证明只有大赦国际才造成变化。 更多的合并拉力请求也并不能确定客户得到的数值是它的两倍。 该版单独引用了AI协助更改较大，agent创建的拉力请求等待审查的时间要长得多，说明了代码生成的速度如何能将工作转移到交付过程的其余部分。
一个顾客的例子使这种区别变得具体。 LinearB 称 Kraken使用其带有AI的MCP服务器来调查审查拖延，发现时区和审查者可用性比大变动更重要。 据报道，Kraken在90%的审阅时间在四个月中减少了一半以上。 公司与Kraken的公开讨论提供了进一步的背景，说明如何利用储存库数据和交付指标来发现瓶颈。
从衡量AI 收养到改进执行
线性B计划为审查，合并和治理等下游工作增加agent技能。 这些是未来的方向，有别于工作室目前的可用性和工作技能的早期进入状态。
更广义的命题是，编码agents需要了解围绕代码的组织：重要的是什么，谁拥有它，在较早的更改失败的地方，哪些限制适用。 Agentic Studio 带来向开发者-生产力平台的命题，其价值最终以交付的修改的质量和流量而不是生成代码的数量来衡量。
本文由 AI图鉴自动抓取并翻译，版权归原作者所有；如需查看原文请点下方链接。
AI Models & Platforms
LinearB Launches Agentic Studio to Give Coding Agents Engineering Context
Add Unite.AI to your preferred sources on Google
LinearB has launched Agentic Studio to turn software-delivery data into context that engineering leaders and coding agents can use to make decisions. Announced November 2, the platform combines natural-language analysis with an engineering context layer spanning the tools teams use to plan, write, review and release software.
The launch reflects a growing problem in AI-assisted development: producing more code does not automatically help an organization deliver useful software faster. Reviews can become congested, changes can grow harder to assess, and agents can begin a task without knowing which services are fragile or why a particular fix matters.
Agentic Studio is available to all LinearB customers, according to the announcement. Agent skills, including an initial capability for matching an agent’s effort to a task, are in early access.
Turning Delivery Data Into Engineering Context
LinearB collects information across software delivery and normalizes it into a common model. The company says it reconciles engineers’ identities across Git and project-management systems, connects changes from ticket to release, and adds AI attribution and benchmarks to help leaders interpret the results.
The important step is connecting those records. A pull request can look straightforward in isolation while affecting a service with recent incidents, frequent rework or complicated ownership. Linking the change to that history gives a reviewer or coding agent information that a repository snapshot alone may not provide.
In the new studio, leaders can ask questions in plain language, obtain suggested next steps and use AI assistance to create metrics, dashboards and other artifacts. LinearB says organizations can define the context available to users with role-based access built into the platform.
That direction aligns with the company’s context-engineering guidance, which emphasizes selecting relevant information rather than filling an agent’s context window indiscriminately. Its framework considers the code, the intent behind a change, connections across systems and the boundaries of what an agent may access.
An Early-Access Skill for Matching Effort to Risk
The first agent skill is designed to decide how much effort a coding task warrants. LinearB says it considers business relevance, complexity and risk, using signals such as rework rates, recent service incidents and ownership history.
A low-risk change can receive a lighter treatment with fewer tokens, while a more consequential change can receive more deliberate analysis and testing. Engineering leaders can see how frequently tasks are routed to low, medium or high effort. The company says the skill works with any coding agent, although the announcement does not provide comparative results across specific agents or quantify realized token savings.
The engineering idea is useful even before a cost claim is established. A uniform reasoning budget treats a minor maintenance change and a sensitive production fix alike. Task-specific context could help allocate more attention where mistakes are expensive, while reducing unnecessary work on simpler changes.
MCP Connects the Data to Existing AI Workflows
LinearB also exposes engineering information through the Model Context Protocol, or MCP. Its MCP product page describes natural-language access to delivery metrics, custom reporting and bottleneck analysis, alongside recommendations such as review automation and routing.
This gives teams a way to investigate engineering data from an AI client instead of manually moving between dashboards. MCP provides the connection to tools and information; the context layer supplies the delivery-specific relationships that make a question about a repository or team meaningful.
There is a distinction to resolve as the new platform rolls out. While the launch describes role-based access and enterprise guardrails, LinearB’s public MCP FAQ still lists permissions-scoped MCP access as a roadmap item. The new studio’s access-control statement should therefore not be read as confirmation that every existing MCP client already inherits every LinearB permission.
Why More AI Output Can Expose Downstream Bottlenecks
LinearB’s announcement cites 2026 AI benchmarks drawn from 2.7 million pull requests and 83,000 developers across 253 organizations. Developers using AI on at least 75% of coding days merged 96% more pull requests per week than a year earlier, according to the company.
Those figures describe observed activity, rather than proving that AI alone caused the change. More merged pull requests also does not establish that customers received twice as much value. The release separately cites findings that AI-assisted changes were larger and agent-created pull requests waited substantially longer for review, illustrating how faster code generation can shift work onto the rest of the delivery process.
A customer example makes that distinction concrete. LinearB says Kraken used its MCP server with AI to investigate review delays and found that time zones and reviewer availability mattered more than large changes. Kraken reportedly reduced review time at the 90th percentile by more than half over four months. The company’s published discussion with Kraken provides further context on using repository data and delivery metrics to uncover bottlenecks.
From Measuring AI Adoption to Improving Delivery
LinearB plans additional agent skills for downstream work such as review, merge and governance. Those are future directions, distinct from the studio’s current availability and the effort skill’s early-access status.
The broader proposition is that coding agents need an understanding of the organization around the code: what matters, who owns it, where earlier changes failed and which constraints apply. Agentic Studio brings that proposition into the developer-productivity platform, with its value ultimately measured by the quality and flow of delivered changes rather than the amount of code generated.
Auto-fetched and translated by AI Atlas. All rights belong to the original author.