Human Review and Compliance Patterns

Organizations with change-management controls, including controls used in SOC 2 programs, may require a human creator or sponsor and an independent human reviewer for code changes. The applicable requirement comes from the organization’s policy; using gh-aw or a particular authentication method does not by itself establish compliance.

An agentic workflow can research a problem and publish its findings and proposed implementation through create-issue. A human reviews the proposal and assigns the issue to a coding agent. The coding agent creates a pull request, which a separate human reviews under the repository’s change-management rules.

This pattern records the human decision to initiate implementation separately from the automated research. Keep the issue, assignment, pull request, and review linked so reviewers can trace the proposal to the resulting change. When human assignment is a required control, the research workflow should not automatically assign the issue or create the implementation pull request.

Research and implementation can use different environments: for example, the research workflow can run on Ubuntu while the coding agent uses a configured Windows environment. The issue carries the findings and acceptance criteria between them. See ResearchPlanAssignOps for the broader research and implementation pattern.

A workflow using create-pull-request can publish through the default GitHub Actions token, a GitHub App, or a configured personal access token. These choices affect the publishing identity. An automation identity does not identify a human sponsor by itself.

Where organizational policy permits it, a personal access token can associate publication with a designated human account. That attribution is not evidence that the person reviewed or approved the generated code. Record the sponsoring decision and require independent review through the repository’s rulesets or branch protection. Follow the organization’s credential ownership, scope, and rotation requirements when choosing this approach.

Useful evidence includes the triggering event and workflow revision, the research issue or change request, the human assignment or sponsoring decision, the proposed diff, required checks, review approvals, and the merge record. Retain workflow artifacts according to the organization’s evidence policy; artifact expiration can otherwise remove supporting execution details.

The distinction between publishing identity, human sponsorship, and code review should be explicit in the control description. See Authentication for supported credentials and Auditing Workflows for run evidence.