Humans on the loop is a position for people in agent-driven software development, named by Kief Morris in Humans and Agents in Software Engineering Loops, in which humans build and manage the working loop instead of either leaving agents to it or micromanaging what they produce. It is defined against two alternatives within the author’s model of a why loop (turning ideas into working software, run by humans) and a how loop (building the software through intermediate artefacts such as specs, code and tests): humans outside the loop, who leave the how loop to agents, and humans in the loop, who act as gatekeepers inside its innermost level, for instance by inspecting each line of generated code. On the loop, humans shape the how loop itself through the agent’s harness — the specifications, quality checks and workflow guidance that control its levels.
Usage
The article identifies the practice of building and maintaining those harnesses, Harness Engineering, as how humans work on the loop. Its clearest test is what happens when an agent’s output, including an intermediate artefact, is not satisfactory: the in-the-loop response is to fix the artefact, by editing it or telling the agent what to correct; the on-the-loop response is to change the harness that produced it so that it produces the wanted results. Outcomes then improve by continuously improving the harness. The article notes that something like this concept has also been described as the “middle loop” — moving human attention to a loop above the coding loop.
When It Applies
The proposal assumes agents that can be given guidance and quality checks to judge their own work, on the “shift left” reasoning that agents, like developers, do better when they can gauge quality themselves than when they rely on someone to check for them. It is offered against the bottleneck created when humans inspect everything agents generate, and against the loss of the internal quality that affects the time and cost of agent work when humans stay out entirely. It is one author’s framing, not a measured result; he extends it to the Agentic flywheel, in which agents themselves recommend and apply harness improvements.