Sheridan — Human Supervisory Control

Author	Thomas B. Sheridan
Work	Telerobotics, Automation, and Human Supervisory Control
Year	1992 (framework roots in Sheridan & Verplank, 1978)
Venue	MIT Press
Type	Book
AM axis	Autonomy / leash
Tags	supervisory-control, levels-of-automation, delegation, human-machine
Actionable	4 / 5

Sheridan gave human factors its vocabulary for degrees of automation. His “supervisory control” describes exactly the relationship a developer now has with a coding agent: a human intermittently instructs a semi-autonomous system, which then closes its own control loop, and monitors the result.

Summary

• Supervisory control, defined. The human does not act directly on the process. Instead they intermittently program and receive information from a computer that itself closes an autonomous loop through its own sensors and effectors. The human sets goals and constraints; the machine executes and reports back.
• Automation is a spectrum, not a switch. The Sheridan–Verplank levels of automation run from the computer offering no assistance, through it suggesting options, narrowing them, executing with approval, executing then informing, up to acting fully autonomously and telling the human only if it decides to. Every real system sits somewhere on this scale — and can sit at different points for different sub-tasks.
• The level determines the human’s cognitive job. Where a task sits on the scale decides what the human must do: choose among options, approve an action, or merely be informed after the fact. Picking the wrong level is a design error, not just a preference.

The one reusable technique

Set the level of automation explicitly, per decision, by risk. Rather than “is this agent autonomous or not,” ask: for this action, should the agent propose and wait for approval, act-then-report, or act silently? High-blast-radius actions (a force-push, a schema migration, a spend) sit low on the scale — propose and wait. Reversible, low-stakes actions can sit high. The scale gives you the words to make that choice deliberately.

Representative quote

Supervisory control means that “one or more human operators are intermittently programming and continually receiving information from a computer that itself closes an autonomous control loop.”

— paraphrasing Sheridan’s definition of human supervisory control (1992)

Paraphrase

The line above is a paraphrase of Sheridan’s definition, not a verbatim quotation — marked as such deliberately. The levels-of-automation scale it references is genuine (Sheridan & Verplank, 1978).

What this means for agentic engineering

Sheridan is the theory behind AM’s Autonomy / leash axis. “Leash length” is not a metaphor invented for agents — it is the levels-of-automation scale applied to a coding agent:

• Name the level for each action class. A mature setup does not run every agent action at the same autonomy level. Code edits inside a worktree can be act-then-report; anything that leaves the machine or is hard to undo drops to propose-and-approve. This is the human-in / on / off-the-loop spectrum made concrete.
• Higher levels shift the human from actor to auditor. As you move up the scale, you stop doing and start judging — which is exactly where Bainbridge’s vigilance problem bites. Sheridan tells you where you are on the delegation scale; Bainbridge warns what it costs you when you climb it without keeping the human engaged.
• Leash length is earned and local. The scale is per-decision and per-risk, never a global “trusted / not trusted” flag. In AM terms, a longer leash is earned through proven verification, granted for a specific action class, and revisited as risk changes.

Together with Bainbridge, Sheridan supplies the two halves of the autonomy axis: how much to delegate (the scale) and what it costs the supervising human (the ironies).