GitHub's July 23 public preview for agent automation controls in Issues is small in scope and large in signal. Agent automations can label, type, assign, and close issues while showing the reason behind each change. Teams can also hold suggested changes for human review instead of applying everything immediately.
The feature adds three concrete control mechanics: approvals, confidence, and rationale. High-confidence actions can be applied automatically, while medium- and low-confidence actions can be held as suggestions. Reviewers can accept or decline individual suggestions, and GitHub exposes a has:suggestions search qualifier so pending agent work can be found. Supported actions include labels, fields, type, close, and assignees across GitHub Agentic Workflows, Copilot cloud agent automations, REST, and GraphQL.
The most important sentence in the source is the caveat: approvals are a workflow convenience, not a security control. An agent with permission to modify issues can still apply changes directly if the system allows it. That distinction is exactly where production AI programs usually get sloppy.
Grey Haven's read: agent governance will often arrive as boring workflow plumbing before it arrives as a grand orchestration layer. Confidence, rationale, and review thresholds inside the queue where work already happens are more deployable than a separate dashboard that operators must remember to check.
The operator heuristic is simple. Use these controls to reduce review load, not to replace permission design. If an agent can close, assign, or relabel operational work, decide which actions are suggestions, which are automatic, who audits the rationale, and what permissions remain outside the agent's reach. Source: GitHub Changelog.
