Human-in-the-Loop: The Feedback Step That Makes It CANI

Input, Transformation, Output — that's a line. It runs once and stops. What turns a line into a loop is this final stage: the human, you, the Architect, closing the circuit. Human-in-the-loop is the difference between an automation that does the same thing forever and a system that gets sharper every cycle. It's also the stage that keeps you in command of an org made of agents — your hand on the wheel, not on the work. This is where FAST becomes CANI: constant and never-ending improvement.

What Human-in-the-Loop actually is

It's three moves, and none of them is "do the work":

• Review — spot-check the output. Not every item, not forever, but enough to know the system is holding its standard. You're the quality bar, sampled.
• Correct — when the agent drifts, you don't redo the task; you fix the system. A wrong output is a bug in the Skill, the input, or the Tools — patch the cause, not the symptom.
• Raise the bar — when the loop is solid, you push it. Tighten the standard, hand it a harder job, widen its scope. The ceiling moves because you move it.

The trap to avoid is collapsing back into the work. The moment you start doing the output instead of steering it, you've quit being the Architect and rehired yourself as the operator. Human-in-the-loop means hand on the wheel — never back in the engine.

How an Architect runs it

You design the feedback, not just the task:

• Decide what needs eyes. High-stakes or low-confidence outputs get reviewed; routine ones get sampled. You calibrate trust to risk.
• Feed corrections back into the Skill. Every fix you make should make the next run better. A correction that doesn't update the system is wasted.
• Promote the loop. As trust earns out, widen the agent's autonomy. The goal is a loop that needs you less over time, not one you babysit forever.

Worked example

The support loop has run for two weeks. You review a sample each morning. You notice the agent keeps over-apologizing on billing questions — technically fine, but off-brand. You don't rewrite those replies by hand. You add one line to the Skill: "On billing, be direct and confident, not apologetic," with an example.

The next day, every billing reply is on-voice. One correction, applied to the system, improved every future output at once. That's CANI: the loop didn't just run again — it ran better, because the Architect closed the circuit. Do that across every pillar and every loop, and you don't have automations. You have an org chart that compounds.