From Assistance to Agency: Rethinking Autonomy and Control in CI/CD Pipelines

This research paper addresses the emerging challenge of designing safe AI agents for CI/CD pipelines by introducing a framework distinguishing between data-plane authority (localized interventions) and control-plane authority (configuration changes). The authors argue that current systems prioritize bounded autonomy with external governance rather than intrinsic safety guarantees, identifying control-plane safety and formalization of autonomy boundaries as critical research gaps.

The integration of AI agents into CI/CD workflows represents a fundamental shift in software development infrastructure, yet lacks standardized frameworks for managing the delegation of decision-making authority. This research tackles a practical problem facing organizations deploying autonomous systems in production environments: how to progressively grant AI agents more operational control while maintaining safety and human oversight.

The distinction between data-plane and control-plane authority is particularly valuable. Data-plane operations like patch generation or test reruns are relatively low-risk and reversible, whereas control-plane decisions affecting deployment policies or approval gates carry significantly higher stakes. Current industry practice concentrates agent autonomy at the data plane, relying on surrounding governance infrastructure rather than building safety directly into agent behavior. This approach reflects a pragmatic but potentially fragile strategy that depends heavily on operational discipline.

The widening gap between deployment momentum and evaluation methodology presents the most concerning finding. Organizations are adopting agentic CI/CD systems faster than the research community can validate their safety properties and develop appropriate assessment frameworks. This creates technical debt in automation governance, similar to regulatory lag in emerging technology sectors.

For software development organizations and DevOps teams, this research signals that investing in explicit governance mechanisms, clear autonomy boundaries, and human-agent coordination protocols will become competitive advantages. The identified research agenda suggests that standardized evaluation frameworks and formalized autonomy specifications will likely emerge within 18-24 months as industry maturity increases. Development teams should audit their current CI/CD authority delegation practices and establish clear recourse mechanisms before expanding agent autonomy.

• →AI agents in CI/CD currently operate primarily at the data plane with constrained autonomy, avoiding high-risk control-plane decisions.
• →Current safety depends on external governance infrastructure rather than intrinsic agent guarantees, creating potential vulnerabilities.
• →Control-plane safety and formalization of autonomy boundaries represent the most urgent open research problems.
• →A significant gap exists between deployment velocity and evaluation methodology for agentic CI/CD systems.
• →Organizations need explicit frameworks for authority transfer and human-agent coordination as autonomous systems expand.