The new human agent working model

Duet Autopilot, Decagon’s self improving agent, now makes more than half of all Agent Operating Procedure (AOP) and tool edits across Decagon deployments, and that share is growing. Our Duetbench-2 benchmark gives us a way to validate the impact of those edits with rigor. This white paper looks at a different question: how does the CX organization's shape, roles, and staffing model change once the agent starts building the agent?

The workstreams Duet is absorbing are the manual, high-volume ones: conversation review and QA, AOP writing and editing, tools editing, and metrics review. In their place, humans can turn more attention to calibration and review, technical governance, prioritization of proposed fixes, and ownership of the knowledge, business logic, and API foundations Duet builds on.

This restructuring is already underway. Customers running Decagon have reorganized into tiger-team models, democratized who can build AOPs and tools, and shifted QA away from creating AI outputs toward reviewing and tuning them. Self-improvement is accelerating a trend already in motion, not starting a new one.

The following pages describe a notable shift playing out across every role in the CX organization. Duet Autopilot absorbs the manual work of building and iterating on the agent, and human roles move toward governance, logic definition, and review. Use this framework to redesign roles, invest in the right foundations, and govern the agent to maximize agent quality as it takes on more of the build.