GitLab 发布 DevSecOps 平台 19.4 版本，为 GitLab Duo CLI 新增了目标驱动模式、三个托管开源权重模型、更多 Model Context Protocol (MCP) 工具，以及更精细的 AI 开销控制能力。此次发布让 GitLab 的智能体不再局限于单个提示词和预定义流程，同时智能体的操作仍受项目现有权限与审批规则约束。
GitLab Duo CLI 中新增的 /goal 命令目前处于公开测试阶段，它接受一个开放式目标，并持续执行任务直到完成任务或达到迭代上限。一个模型负责实施变更，另一个模型对照既定目标检查进度。开发者可以停止运行、更改目标或重新引导智能体。任务在本地运行，并保留项目现有的上下文和控制机制。
“此次发布将智能体自动化从单个开发者工具变成企业能够在已有控制体系下快速扩展的能力。”
GitLab 这里使用的“验证”一词有特定含义。第二个模型判断结果是否满足自然语言描述的目标，并非对代码进行形式化验证。该命令会报告无法解决的任务，迭代上限可避免任务无限执行。因此清晰的验收标准依然十分重要，尽管开发者无需监督每一个中间步骤。
GitLab Duo Agent Platform 还新增了对托管开源权重模型 Kimi K3、MiniMax M3 和 GLM 5.3 的支持。GitLab 表示，与许多同类前沿模型相比，这些模型的每 GitLab 积分最多可支持 4 倍的调用量。群组所有者可以为某项功能设置默认模型，并限制团队可选用的模型。企业可在常规工作中选用低成本模型，把成本更高的模型留给更适合它的任务。公告中未提供各模型的基准测试结果或每次调用的绝对价格，因此客户仍需针对自己的工作负载测试这些选项。
“运行自动化的平台决定了智能体可以使用哪些工具，并统计消耗的资源。”
扩展后的 GitLab MCP 服务器工具让第三方客户端中的智能体能够触发和检查流水线、打开和审查合并请求、更新工作项、浏览代码库以及管理漏洞。只读工具默认设置为“始终允许”，写入和删除操作默认设置为“始终询问”。管理员可以在群组或项目级别更改这些规则。大多数 MCP 工具在免费版、高级版和旗舰版可用，但漏洞管理工具仅旗舰版支持。
这是让编码智能体通过 MCP 执行操作而不授予无限制访问权限的更大规划的一部分。InfoQ 在 2025 年报道称，谷歌面向 Android Studio 的 Agent Mode 可以使用 MCP 访问外部工具。GitHub Copilot 也在其主要界面支持 MCP，并提供用于控制服务器访问权限的企业策略。GitLab 的方法是对每个工具单独设置权限策略，默认在修改操作前设置人工审核检查点。
当智能体自动运行而不是对单个提示词做出响应时，成本控制就变得更加重要。DevOps 咨询公司兼 GitLab 合作伙伴 ReleaseTEAM警告称，“实际的风险在于自动化”，因为与每个流水线故障或新合并请求相关的流程每天可能运行数百次。该公司建议在将流程连接到高频触发器之前设置用量上限，再通过小规模试点监测资源消耗。
第四项主要更新是成本归属统计。现在的积分使用导出文件为每一条计费事件单独生成一行记录，包含用户、项目、会话、流程类型、积分数和 Token 数量。新的设置页面让所有者可以设置默认单用户用量上限，还能单独覆盖个别用户限额，无需调用 GraphQL API。开发者可以查看自己的消耗情况，当按需用量达到月度上限的 50% 或 80% 时，账单经理会收到预警。
本次更新是 GitLab 在 2026 年对 GitLab Duo 一系列迭代的延续。GitLab 19.2 让智能体处理依赖项修复和安全评审，但仍需人工审批变更。更早之前，18.10 和 18.11 版本引入了固定费率的自动化代码评审、Duo Agent Platform 的免费层级访问以及月度积分上限。19.4 版本将这些能力扩展到运行时间更长的开发任务，并让管理员更清晰地了解由此产生的成本。
GitLab 19.4 于 2026 年 9 月 17 日发布。/goal 命令面向 GitLab.com、GitLab 自托管版本和 GitLab Dedicated 的 Premium 和 Ultimate 客户，要求 GitLab 19.3 及以上版本搭配 Duo CLI 9.17.0 或更高版本。该命令与增强版 MCP 工具仍处于公开测试阶段，新的积分用量明细功能已正式可用。
查看英文原文：https://www.infoq.com/news/2026/10/gitlab-19-4-goal-driven-agents/
本文由 AI宝库自动抓取并翻译，版权归原作者所有；如需查看原文请点下方链接。
The GitLab publish DevSecOps platform, version 19.4, added a target driver model for GitLab Duo CLI, three hosting open-source weight models, additional ModelContext Protocol (MCP) tools and a more sophisticated AI cost control capability. This time publish made Gitlab ' s AI agent no longer limited to a single prompt and a predefined process, while AI agent ' s operation remained subject to the existing permission and approval rules of the project.
The new /goal command in GitLab Duo CLI is currently in the public testing phase, accepting an open target and continuing to perform its tasks until the completion of the mission or reaching an iterative ceiling. One model is responsible for implementing changes, while the other examines progress against established targets. Developer can stop, change or redirect AI agent. The mission operates locally and maintains the existing context and control mechanisms of the project.
“This time publish automates AI agent from a single developer tool to an enterprise's ability to expand rapidly under an existing control system.”
GitLab here uses the word “certification” in a specific sense. The second model judged whether the results met the objectives of the natural language description and did not formally validate the code. The order reports on tasks that cannot be resolved, and an iterative ceiling avoids unlimited implementation. Clear acceptance criteria therefore remains important, although developers need not monitor every intermediate step.
GitLab Duo Agent Platform has also added support for hosting open source weight models Kimi K3, Mini Max M3 and GM 5.3. GitLab indicates that each GitLab score of these models supports a maximum of four times the call amount compared to many of the same front models. Group owners can set default models for a function and limit the models that teams can choose to use. Enterprises can choose to use low-cost models in their regular work, leaving more cost-effective models to tasks more appropriate to them. The bulletin does not provide the baseline test results of the models or the absolute price of each call, so the client still needs to test these options for his/her work load.
“The platform for operational automation determines which tools AI agent can use and calculates the resources consumed.”
The expanded GitLab MCP server tool allows AI agent in third-party clients to trigger and check water flow lines, open and review merger requests, update work items, browse the code library and manage loopholes. Read-only tool default is " always allowed ", write and delete operation default is " always asked ". Administrators can change these rules at group or project level. Most MCP tools are available in free, advanced and flagship versions, but the gap management tool is supported only by flagship versions.
This is part of a larger plan to allow code AI agent to operate through MCP without granting unrestricted access to permission. InfoQ reported in 2025 that Agent Mode, Google for Android Studio, could use MCP to access external tools. GitHub Copilot also supports MCP in its main interface and provides an enterprise strategy for controlling server access to permission. GitLab uses the permission policy for each tool and defaults to set a manual check point before the modification.
Cost control becomes more important when AI agent functions automatically rather than responding to a single prompt. DevOps Consulting and GitLab Partner ReleaaseTEAM warned that “the real risk lies in automation” because the processes associated with each streaming line failure or new merger request may operate hundreds of times a day. The company proposes to set a maximum usage before connecting the process to a high-frequency trigger, followed by small-scale pilot monitoring of resource consumption.
The fourth major update is cost attribution statistics. The current score export file generates a separate line of records for each fee event, including the number of users, items, sessions, process type, points and Token. The new setup page allows the owner to set the default single-user limit and to cover the individual user limit separately without calling the GramQL API. The developer can check its consumption and the billing manager receives an early warning when it reaches 50 or 80 per cent of the monthly ceiling at the required level.
This update is a series of iterative extensions of Gitlab in 2026. GitLab 19.2 AI agent was asked to process the dependency restoration and security assessment, but manual approval of the change was still required. Earlier, versions 18.10 and 18.11 introduced the automated code evaluation of fixed rates, free level access to Duo Agent Platform and a monthly credit ceiling. 19.4 The version extends these capabilities to longer operational development tasks and provides managers with a clearer picture of the resulting costs.
GitLab 19.4 September 17, 2026 publish. The /goal command is for Premium and Ultimate clients of GitLab.com, GitLab and GitLab Dedicated, requiring GitLab 19.3 and above to match Duo CLI 9.17.0 or higher. The command and the enhanced MCP tool are still in the open-test phase and the new credits are fully available.
See the original English language version at https://www.infoq.com/news/2026/10/gitlab-19-4-goal-driven-agendas/
Auto-fetched and translated by AI Gem Box. All rights belong to the original author.