Agentic Autonomy Levels (addyo.substack.com)

A new model for agentic engineering has been announced, focusing on levels of autonomy that can be directly integrated into software development workflows. The framework categorizes autonomy into two axes: agency, which measures the degree of independence an agent has in decision-making, and orchestration, which assesses how agents are managed and coordinated. As AI tools like Claude Code and Codex evolve, this model highlights the transition from simple prompting to more complex, goal-driven activities where multiple agents work simultaneously on tasks, thereby increasing efficiency and reducing human oversight. This model is significant for the AI/ML community as it lays out a structured approach for leveraging automated agents in software engineering, potentially transforming how developers interact with AI. The six defined levels of autonomy range from basic assistance to full orchestration of numerous agents, enabling more sophisticated task management. Key implications include a clearer understanding of risk and reversibility in deploying AI systems, guiding developers on when to grant more autonomy based on task complexity. As systems advance towards high autonomy, it becomes crucial to establish clear contracts outlining goals, scopes, evidence requirements, and escalation paths to ensure reliable and accountable agent behavior.