GitHub Issues Adds Confidence Thresholds and Per-Item Approvals for Agent Actions, but They Do Not Constitute a Permission Boundary

GitHub is embedding agent governance into the Issues state-change workflow. The new public preview covers approvals, confidence, and rationale: agents can first submit changes to a suggestions panel, where each supported action is labeled with high, medium, or low confidence and accompanied by a rationale. The default logic can apply high-confidence actions immediately while leaving medium- and low-confidence results for users to accept individually or in batches. Repository administrators can also adjust automation thresholds and use `has:suggestions` to find items awaiting review.

The first phase supports adding labels, changing issue types and fields, closing issues, and assigning them to users or agents. These capabilities are available to Copilot cloud agent, GitHub Agentic Workflows, the REST API, and the GraphQL API. After upgrading, existing Agentic Workflows can include issue intent in safe outputs. To require intent data, set `issue-intents: true` in the workflow frontmatter. This gives teams an option for large-scale classification, metadata completion, and spam-issue filtering that falls between fully automated and entirely manual workflows.

Engineering teams should not treat confidence scores as calibrated risk probabilities. GitHub has not disclosed how the scores are generated, the error rates at each threshold, or cross-repository evaluations. Agent-provided rationales are also only audit clues; they do not necessarily faithfully represent the model’s internal reasoning. More importantly, approvals are a workflow convenience rather than a security control: an agent that already has write access to issues can still bypass suggestion mode and make changes directly. Deployments should preserve least-privilege access, separately log API activity and audit events, and recalibrate auto-apply thresholds using real-world data on incorrect labels and erroneous closures.