Skip to main content

Mechanical Feedback Loops

This method turns work into control evidence.

A loop is not closed because work ended. It closes when the result changes the next pass. The job is to make that change mechanical enough that another agent can read it.

Inputs​

  • A completed or recurring piece of work.
  • A declared value setpoint: what good the loop serves.
  • A belief horizon: what was assumed true enough to act.
  • A gauge that can contradict the claim.
  • A lever that can change the next run.
  • A state store where proof can live.

Method​

  1. Setpoint. Say what the loop tries to hold or raise: clarity, trust, quality, speed, safety, revenue, learning, or confidence.
  2. Reality read. Record the closest ground source, then read the gauge that could prove the claim wrong.
  3. Variance. State the gap between the setpoint and the gauge reading.
  4. Correction. Change the lever that naturally touches the next run: rule, skill, route, demand, deletion, prompt, checklist, test, or status reader.
  5. Proof and next question. Leave the smallest artifact another agent can consume, name the baseline lift, and feed the sharpest unanswered question into the next loop.

Controller Shape​

Use this compact close when a loop matters. It has eight fields because controller proof must preserve the whole path from intended setpoint to next question.

Setpoint:
Ground:
Gauge:
Variance:
Correction:
Proof:
Baseline lift:
Next question:

Then run the reflection close:

  1. Are we doing the right thing?
  2. Is the endeavour meaningful and aimed at the greater good?
  3. Are we pursuing it as effectively as possible?

Checks​

  • The setpoint serves a real value, not a proxy alone.
  • The belief horizon is visible enough to challenge.
  • The gauge can go red.
  • The correction changes a future default.
  • The proof artifact survives outside the session.
  • The next question is more precise than the starting question.

Failure Modes​

  • Event thinking — treating one bad result as isolated instead of asking what kept producing it.
  • Gauge without lever — measuring drift without installing a correction.
  • Lever without value — making action easier before deciding whether the action is good.
  • Reflection theater — writing a lesson that no future loop reads.
  • Automation before validation — speeding up a loop before proving it is worth running.
  • Fixed setpoint forever — holding a target that reality has shown should rise or die.

Proof Of Done​

The loop is closed when the proof artifact lets the next operator answer:

  • What was the setpoint?
  • What did reality show?
  • What changed because of the variance?
  • Where will the next loop read that change?
  • What question should it answer first?

Context​

  • depends-on Control System — values and beliefs become gauges, levers, and correction.
  • depends-on Virtuous Feedback Loop — validation plus virtue plus feedback plus compounding.
  • pairs-with Agentic Prompt Loops — agent loops need mechanical closes.
  • proved-by Loop Registry — recurring loops need a status reader and kill signal.
  • applies-to Delegation — return loops prove delegated work stayed inside intent.

Questions​

What changed in the system that makes the next good action easier by default?

  • Which gauge contradicted the starting belief?
  • Which correction will the next agent naturally hit?
  • Which loop should be killed because it no longer serves the setpoint?