Skip to main content

Loops

What loop is already training the next result?

A loop is what happens when the result of one action shapes the next action.

That is all. You do something, reality answers, and the answer changes what you do next.

Part of the Flow KB — the Mechanism facet: AI amplifies whichever loop you feed it.

Why Loops Matter​

Most people do not think in feedback loops. They think in events.

Event thinking says: "That went badly."

Loop thinking asks: "What kept producing that result?"

The difference matters because AI makes loops faster. If you feed an AI tool rushed prompts, vague goals, and untested assumptions, it gives you more of that pattern. If you feed it clear intent, useful context, measured outcomes, and better questions, it helps you improve the pattern.

AI amplifies the loop you already run.

The Everyday Version​

You already know loops from normal life.

A child tries a joke. People laugh. The child tells more jokes like it.

A manager only rewards urgent work. The team creates more urgency.

A person expects a hard conversation to go badly, enters guarded, gets a guarded response, then treats that response as proof they were right.

In each case, the output returns as input. The loop trains the next move.

The Thought Loop​

A thought loop is the human version of a feedback loop.

It runs like this:

  1. Experience — something happens.
  2. Intent — you decide what you want to do with it.
  3. Decision — you choose how to act.
  4. Capability — you use or improve the skill needed to act.
  5. Outcome — reality answers.
  6. Expectation check — you ask, "Is this what I expected?"
  7. Adjustment — if not, ask why and change the loop.
  8. Amplification — if yes, ask how to get more of the good result.
  9. Better question — turn the lesson into a sharper idea for the next loop.

The loop closes when the outcome changes the next question.

Where AI Fits​

AI enters the loop as an amplifier, not a substitute for judgment.

It can help you notice patterns, compare options, draft actions, simulate outcomes, remember decisions, and test whether the result matched the intention.

It cannot choose your values for you. It cannot decide what a good life, good company, or good outcome should mean. Those are setpoints. A setpoint is the thing the loop is trying to move toward.

Bad setpoint, faster loop: faster drift.

Better setpoint, measured loop: compounding agency.

Use It​

Before choosing an AI tool, name the loop.

What is happening now? Name the repeated pattern, not only the latest event.

What do you want more of? Name the setpoint: clarity, trust, speed, quality, calm, revenue, learning, or confidence.

What action will test it? Choose the smallest move that can change the pattern.

What will you measure? Decide what outcome would prove the loop improved.

What question comes next? Write the question that feeds the next idea into the hopper.

Agent Loops​

Prompting is the manual version. Loop engineering is the system version.

A prompt asks an agent to take one guided pass. A loop gives the agent a repeatable contract: when to start, what context to load, what tools it may use, what proof it must leave, and when to stop.

A working coding-agent loop has six parts:

  1. Trigger or cadence — the event, command, schedule, or goal that starts the loop.
  2. Isolated workspace — the branch, worktree, sandbox, or permission mode that bounds change.
  3. Skill and context — the procedure and state the agent loads before acting.
  4. Connectors and tools — the files, APIs, agents, browsers, CLIs, or message buses it can reach.
  5. Verifier — the check that proves the next pass is earned.
  6. External state — the snapshot, issue, receipt, plan, or database record that survives chat context.

Good loops also state the guardrails before the first run:

  • Stop condition: what exact proof ends the loop?
  • Token budget: how many turns, model calls, or tokens can it spend?
  • Escalation rule: when does it pause for a human?
  • Proof record: where does the loop write what happened?

Use agent loops for recurring discovery, triage, verification, and bounded fixes. They are useful when the pattern repeats and the success signal is checkable.

Avoid agent loops when the values are unclear, the change is high risk, the verifier is vague, or the work still needs direct human judgment. Faster drift is still drift.

Loop Design Checklist​

  • Name the purpose.
  • Name the trigger or cadence.
  • Bound the workspace and permissions.
  • Load the skill and state store.
  • List the tools and connectors.
  • Define the verifier.
  • Define the stop condition.
  • Set the token and turn budget.
  • Set the escalation rule.
  • Name the proof artifact.

Build the loop. Stay the engineer.

Example​

Weak loop:

  • Experience: the team keeps missing deadlines.
  • Reaction: ask AI to write a stricter reminder.
  • Outcome: people comply for a week, then drift.
  • Question missed: why does the deadline keep becoming invisible?

Better loop:

  • Experience: deadlines keep slipping.
  • Intent: make commitment visible before work starts.
  • Decision: ask AI to turn each task into owner, date, risk, and proof.
  • Capability: practice smaller commitments.
  • Outcome: fewer vague tasks enter the week.
  • Expectation check: did fewer tasks slip?
  • Better question: which commitment still lacked proof?

The second loop improves the system. The first loop only pushes harder on the symptom.

Context​

  • AI — the wider guide to using AI to amplify agency.
  • AI Toolkit — choose practical AI tools after naming the loop.
  • Skills — turn repeated agent work into reusable operating context.
  • Context Graphs — how decisions become memory for the next loop.
  • See the System — draw signals, gauges, and controls before building.
  • Performance — measure whether the loop improved reality.

Questions​

Which repeated pattern are you treating as a one-off event?

  • What result keeps coming back as input?
  • What outcome would prove the loop improved?
  • What better question should go back into the ideas hopper?