GitHub Issues adds AI agent approval workflow and confidence scoring

GitHub introduces granular controls for AI agent automations in Issues

GitHub has released a new set of controls that let developers decide which automated issue management actions take effect immediately and which require human review. The features, now in public preview, address a core tension in AI-assisted development: balancing automation efficiency with the need for oversight on critical decisions.

According to GitHub's announcement, the platform now provides three interconnected capabilities for managing agent-driven changes to issues: approval workflows, confidence scoring, and action rationale.

How the new controls work

The approval system allows teams to configure automations to suggest changes rather than apply them directly. When an agent proposes a change—such as labeling an issue, assigning it to a team member, or closing it—the suggestion appears in a panel on the issue page. Developers can accept or decline individual suggestions or process them in bulk.

Confidence scoring adds a layer of intelligence to this workflow. Agents now rate each supported action as high, medium, or low confidence. High-confidence changes apply automatically by default, while medium and low-confidence actions are held for review. Repository administrators can adjust the confidence threshold to match their team's risk tolerance and workflow preferences.

Every automated action now includes rationale—a recorded explanation of why the agent made that decision. This reasoning appears both on automatically applied changes and on suggestions awaiting review, creating an audit trail and helping developers make informed decisions about which suggestions to accept.

Developers can use the search query has:suggestions to find issues with pending changes. GitHub notes that approvals function as a workflow convenience rather than a security boundary—agents with permission to modify issues can still apply changes directly if configured to do so.

Supported actions and implementation

At launch, the controls cover changes to labels, issue type, custom fields, assignees, and issue closure. The features work with GitHub Agentic Workflows and Copilot cloud agent automations, and are accessible through REST and GraphQL APIs.

Teams using GitHub Agentic Workflows need to upgrade their workflows to add issue intent support. After upgrading, supported safe outputs automatically include intent information while maintaining backward compatibility. Specific safe outputs that support intents include set-issue-type, set-issue-field, add-labels, close-issue, assign-to-agent, and assign-to-user.

For Copilot cloud agent automations, no update is required—teams can create an automation from the Automations pane in their repository's Agents tab.

Use cases

GitHub outlines several practical applications. Triage automations can label, type, and prioritize incoming issues with reasoning attached, eliminating the need for separate triage comments. Metadata enrichment can backfill missing information on older issues. Spam detection can flag suspected spam with explanations and hold uncertain cases for human review.

The system's flexibility means a small team moving quickly might allow agents to run autonomously, while a high-traffic public repository can require review for most changes. Because low-confidence actions are held back by default, developers spend time only on changes that genuinely need a second look.

Why it matters

These controls address a fundamental challenge as AI agents take on more repository management tasks: maintaining human judgment where it counts while capturing efficiency gains from automation. By making agent reasoning visible and giving teams granular control over which actions require approval, GitHub is building trust in automated workflows without forcing teams into all-or-nothing adoption. The confidence scoring mechanism is particularly significant—it creates a middle ground between manual triage and full automation, letting teams gradually expand what they delegate as they build confidence in agent performance.

The features are available now in public preview. Details were first reported by GitHub in their changelog.