Supervisory Engineering Work - Annie Vella
Annie Vella argues that AI coding agents have changed software engineering from writing code to supervising it, and that human factors research on supervisory control from the 1970s and 1990s explains what that new job demands. She uses those decades-old frameworks to map today's levels of AI autonomy onto engineers' roles.
- Her Masters research found engineers shifting from creating code to verifying it, through testing and reviewing, with a new activity she calls supervisory engineering work: directing AI, evaluating its output, and correcting it.
- Supervisory control research from Sheridan and Verplank maps closely onto the role: plan and command, monitor, intervene, and calibrate trust.
- Agents hand over whole tasks ("trading") rather than working alongside the engineer ("sharing"), and the level of autonomy is often left to individuals, defaulting to whatever the tool allows.
Annie Vella's core argument is that AI is automating the writing of code, and that this shifts software engineers from creating to verifying: directing AI, evaluating what it produces, and correcting it when it's wrong. She calls this "supervisory engineering work." Her claim is that this is not just a new job with a new name but a materially different role, one that needs its own skills, training, hiring criteria, and definitions of good performance. She pushes back on the idea that AI simply means more demand for engineers under Jevons Paradox. More demand may be true, but the day-to-day activities are changing, and the industry hasn't caught up with what it means to hold engineers accountable for code they didn't write or even read.
Her main contribution is a historical frame. Vella points to human factors research, which began in WWII when aircraft outpaced pilots and "pilot error" gave way to the recognition of design error, and then to supervisory control, developed largely by Thomas Sheridan at MIT from the late 1960s. Before AI, engineers worked in something like direct control, closing the feedback loop themselves through interpreting requirements, editing code, and reacting to results. Coding agents now close that loop on their own, writing, testing, and fixing, which moves the human into a supervisory role. She draws a useful distinction between "sharing" the load, as with early copilot tools, and "trading" it, where whole tasks are handed off and not always taken back. She maps Sheridan and Verplank's supervisor roles (plan, command, monitor, intervene, trust) onto her research findings, with a table lining up each role with directing, evaluating, correcting, and calibrating trust. The 1978 Sheridan and Verplank report on undersea teleoperators, with its ten levels of automation in human-computer decision-making, is the source she treats as a goldmine of applicable research.
The piece is a companion to a keynote at Agentic Engineering Day in Auckland on 22 October 2026, and it's best read as a framing essay rather than a set of prescriptions. Its value for a read-it-later queue is the argument that the field should borrow decades of research on human-automation interaction instead of reinventing answers about trust, intervention, and accountability. Vella doesn't resolve those questions here. She mostly makes the case that they're the right questions to be asking.