GitHub Issues 中的 Agent 自动化控制功能进入公开预览

Agent 自动化越来越多地为你完成 Issue 标签标记、类型设置、指派和关闭等操作。GitHub Issues 现在会展示每项变更背后的原因，并允许你在变更生效前进行审核，因此你可以自行决定自动化程度，以及何时保留人工参与。

我们发布了三项新能力：

• **审批：**提示自动化将变更作为建议提出，而不是直接应用；这些变更将等待在 Issue 的面板中，而不会立即生效。你可以逐项接受或拒绝，也可以一次性全部接受或拒绝。
• **置信度：**Agent 会将每项受支持操作评为高、中或低置信度。高置信度变更会自动应用。GitHub Issues 会将中、低置信度变更保留为建议，供你审核。
• **理由说明：**每项受支持操作都会记录其背后的原因，无论该操作是自动应用还是等待审核。你将获得关于变更内容及其原因的审计记录，并可在作出决定前查看每项建议的推理依据。

使用 Issue 搜索中的 has:suggestions，可查找存在待审核变更的 Issue。仓库管理员还可以配置自动化级别，以设置置信度阈值，控制哪些变更自动应用、哪些变更保留等待审核。

审核变更是可选的。如果你让自动化直接应用变更，它就会直接应用；如果你让它提出建议，则会增加一个审核步骤。快速推进的小团队可以让 Agent 自主运行，而繁忙的公共仓库则可以将变更保留待审核。由于 GitHub Issues 会暂缓低置信度操作，你只需花时间处理最可能需要再次检查的变更。

审批是一项工作流便利功能，而不是安全控制措施。它不会强制执行服务端边界；拥有修改 Issue 权限的 Agent 可以直接应用变更，而不是将其作为建议提出。

理由说明、置信度和审批功能可与 GitHub Agentic Workflows 和 Copilot 云端 Agent 自动化配合使用，也可通过 REST 和 GraphQL API 使用。发布时，这些功能覆盖对标签、字段、类型、关闭状态和受指派人的变更。

如果你已在使用 GitHub Agentic Workflows，请升级工作流以增加 Issue 意图支持。升级后，受支持的安全输出会自动包含意图信息，同时保持向后兼容。若要针对特定安全输出强制要求意图，请在工作流 frontmatter 中设置 issue-intents: true。

设置 issue-intents: false 可明确选择退出。以下安全输出支持 Issue 意图：set-issue-type、set-issue-field、add-labels、close-issue、assign-to-agent 和 assign-to-user。

无需更新。请在仓库 Agents 选项卡的 Automations 面板中创建自动化，并使用受支持工具查看置信度和理由说明的实际效果。

• **分诊：**自动为新提交的 Issue 添加标签、设置类型并确定优先级。每项操作都会附带其推理依据，而无需新增分诊评论；你可以在变更应用前审核需要审核的操作。
• **元数据补全：**为提交时缺少标签、类型或字段值的 Issue 补充这些信息，并附带理由说明及可选的审核步骤。
• **垃圾信息检测：**标记疑似垃圾信息并说明原因，同时将不确定的案例保留给你审核。

理由说明、置信度和审批现已进入公开预览。立即使用受支持工具创建自动化，体验建议和理由说明功能，或在 GitHub Community 中参与讨论。

Agent automations increasingly label, type, assign, and close issues for you. GitHub Issues now shows the reason behind each change and lets you review them before they’re applied, so you decide how much to automate and when to stay in the loop.

We’ve released three new capabilities:

• Approvals: Prompt your automation to suggest instead of applying, and the changes wait in a panel on the issue instead of taking effect. Accept or decline each one, or accept or decline all at once.
• Confidence: Agents rate each supported action high, medium, or low confidence. High-confidence changes apply automatically. GitHub Issues holds medium and low ones as suggestions for you to review.
• Rationale: Every supported action records the reason behind it, whether it applies automatically or waits for review. You get an audit trail of what changed and why, and can see the reasoning on each suggestion before making a decision.

Use has:suggestions in issue search to find issues with changes waiting for review. Repository admins can also configure the automation level to set the confidence threshold, controlling which changes apply automatically and which are held for review.

Reviewing changes is optional. If you tell your automation to apply directly, it will. If you tell it to suggest, you get a review step. A small team moving fast might let the agent run on its own, while a busy public repository can hold changes for review. Because GitHub Issues hold back low-confidence actions, you only spend time on the changes most likely to need a second look.

Approvals are a workflow convenience, not a security control. They don’t enforce a server-side boundary, and an agent with permission to change issues can directly apply changes rather than suggest them.

Rationale, confidence, and approvals work with GitHub Agentic Workflows and Copilot cloud agent automations, and are also available through the REST and GraphQL APIs. At launch, they cover changes to labels, fields, type, close, and assignees.

If you already use GitHub Agentic Workflows, upgrade your workflows to add issue intent support. After upgrading, supported safe outputs include intent information automatically while remaining backward compatible. To require intents for a specific safe output, set issue-intents: true in your workflow frontmatter.

Set issue-intents: false to explicitly opt out. Issue intents are supported for the set-issue-type, set-issue-field, add-labels, close-issue, assign-to-agent, and assign-to-user safe outputs.

No update is needed. Create an automation from the Automations pane in your repository’s Agents tab and use supported tools to see confidence and rationale in action.

• Triage: Label, type, and prioritize incoming issues automatically. Each action carries its reasoning without adding a triage comment, and you can review the ones you want before they apply.
• Metadata enrichment: Backfill missing labels, types, or field values on issues that were filed without them, with the reasoning attached and an optional review step.
• Spam detection: Flag suspected spam with a reason and hold uncertain cases for you to review.

Rationale, confidence, and approvals are available now in public preview. Create an automation today using the supported tools to see suggestions and rationale in action, or join the discussion within GitHub Community.