Feedback Loops

Intent

A feedback loop occurs when a system's output circles back as input to the next cycle. reinforcing loops amplify (compound interest, viral growth, bank runs, death spirals). balancing loops self-correct (thermostats, price discovery, immune response). the critical complication is delay: when delay is long relative to response time, even well-designed balancing loops produce oscillation and overshoot. operators systematically mismanage these systems because they can't intuit multi-step delayed causation. use this skill when you observe non-linear surprise (collapse, oscillation, death spiral, growth flywheel), you're intervening in a complex system and success depends on how it responds, or trends aren't extrapolating well. do not use when the decision is a one-shot linear choice with no feedback to future decisions, or when an exogenous shock so large it dominates all internal dynamics is the obvious explanation.

composes with: second-order-thinking · s-curve-technology-adoption · prisoners-dilemma · probabilistic-thinking.

Inputs

context required:

• specific system name and the variable you care about tracking (e.g., "product inventory in our supply chain", "team headcount over hiring cycles", "price in this market")
• concrete observed behavior: oscillation, collapse, unexpected growth, or stalled maturity
• rough timeline: how fast is the variable changing, and over what window do you observe the pattern

external connections:

• none required. this skill is structural analysis only. if you're modeling in software, you may feed outputs to a system dynamics simulation tool (stella, vensim, anylogic), but the diagnosis itself requires no external api.

edge cases:

• insufficient data to map loops: hand-waving without a causal structure makes diagnosis impossible. halt and ask for concrete numbers or behaviors.
• decision too time-bounded for delays to matter: if your intervention window is shorter than the known delays in the system, feedback loops may not drive the outcome. clarify the real timeline.
• oscillation blamed on external variability without checking internal loop generators: most operators assume randomness; check internal loops first.

Procedure

Feedback-Loop Diagnosis process:

1. name system and variable of interest. write down the system (e.g., "inventory management in our 3-tier supply chain") and the specific variable you're tracking (e.g., "warehouse stock level in units"). without a named variable, analysis becomes vague narrative.

   • input: case description from user
   • output: one-line system statement + variable name

2. list drivers and outputs. what inputs change your variable? what does your variable change in turn? trace one full cycle. stay concrete (not "management decisions" but "hiring rate in response to queue depth").

   • input: system name + variable
   • output: list of 3-5 direct causal links (A → B, B → C, etc.)

3. identify loops. trace chains where a variable feeds back to itself. most systems have 2-4 dominant loops. draw or write out the full cycle for each.

   • input: drivers and outputs list
   • output: 2-4 loop chains, each written as a causal sequence (e.g., "high stock → longer wait to sell → lower prices → lower production → lower stock → repeat")

4. classify each loop (R or B). count the number of negative causal links in each loop. if even (including zero), it's reinforcing (R). if odd, it's balancing (B). sign-counting is mechanical and removes debate.

   • input: loop chains
   • output: each loop labeled R or B with sign-count shown

5. locate delays. where does a cause take significant time to produce its effect? list the approximate magnitude (hours, weeks, months, years). delays are where human intuition fails and where oscillation is born.

   • input: loop chains
   • output: list of delays with rough magnitude (e.g., "hiring-to-productivity: 3 months", "demand-signal-to-production: 2 weeks")

6. identify dominant loop. in growth phase, reinforcing loops dominate. in maturity, balancing loops catch up. in crisis, suppressed reinforcing loops take over. which loop's behavior matches what you observe?

   • input: observed behavior + classified loops + delays
   • output: one loop named as dominant, with explanation of why it matches observations

7. map stocks and flows. stocks are accumulations (headcount, inventory, cash, customer base). flows are rates (hiring rate, production rate, spend rate, churn rate). a positive flow can leave a stock dangerously low. distinguish them.

   • input: system description
   • output: explicit list of stocks and flows in your system

8. predict behavior pattern without intervention. pure reinforcing loop alone → exponential growth or collapse. pure balancing loop → equilibrium. reinforcing loop plus significant delay → overshoot then oscillation. reinforcing plus balancing competing → s-curve. if prediction mismatches observed behavior, you've missed a loop or misclassified one.

   • input: dominant loop classification + delays + stocks/flows
   • output: predicted behavior pattern and timeframe (e.g., "oscillation with 6-month period, slowly damping")

9. find leverage using Meadows hierarchy. from lowest to highest: parameters (numbers), buffers (capacity), structures (who talks to whom, feedback paths), delays (can they be shortened?), balancing loops (can you weaken one?), reinforcing loops (can you amplify or suppress one?), goals (what are you actually optimizing?), paradigm (what worldview makes this system inevitable?). most failed interventions push parameters; move up.

   • input: system diagnosis + observed behavior
   • output: ranked leverage points with one recommended intervention at the structural level or higher

10. stress-test intervention against system response. balancing loops fight back; reinforcing loops restore trajectory. does your proposed move change structure, or just symptom? what will the system do in response? what's the backfire risk?

    • input: proposed intervention + identified loops
    • output: predicted system response + backfire risk + falsifier (observable that would prove diagnosis wrong)

Decision Points

if user describes a one-shot linear choice with no feedback to future decisions:

• do not activate this skill. redirect to probabilistic-thinking or second-order-thinking. feedback loops require at least two cycles.

if user observes oscillation or collapse but blames external variability (market shock, competitor move, macro event):

• check internal loop generators first. sterman (1989) found that operators in the beer distribution game blamed constant demand variability; the oscillation was self-inflicted. ask: "what part of this pattern could your own system create?" if the shock is truly exogenous and dominates, redirect.

if user has concrete case and wants to run diagnosis immediately (engine mode):

• run the process directly, step-by-step, with their numbers. do not coach.

if user is unfamiliar with feedback loops or has no concrete case yet (coach mode):

• respond one step at a time. after each step, wait for user input before advancing. do not jump ahead or lecture.

if user identifies a balancing loop that is currently being suppressed (e.g., "we ignore customer complaints to hit short-term targets"):

• flag as high-risk. suppressed balancing loops often cause later collapse. leverage point is to restore the signal, not strengthen parameters.

if user's proposed intervention is purely parametric (raise bonus, add metric, hire faster):

• ask: "what loop or delay does this change?" if the answer is "none", the intervention is noise in a structurally-driven system. move up the hierarchy.

Output Contract

deliver a Feedback-Loop Diagnosis document with these sections:

System / variable: <name of system and the specific variable tracked>

Loops identified:
  R1: <causal chain with signs counted>
  B1: <causal chain with signs counted>
  [repeat for each loop]

Delays:
  <location>: <rough magnitude> (e.g., "hiring-to-productivity: 3 months")
  [repeat for each significant delay]

Stocks and flows:
  Stocks: <accumulations in the system>
  Flows: <rates driving those stocks>

Dominant loop: <R or B loop name> ,  matches observed behavior because <explanation>

Predicted behavior without intervention: <pattern and timeframe>

Leverage points (ranked low to high):
  1. parameter: <example>
  2. buffer: <example>
  3. structure: <recommended move>
  4. delay: <if modifiable>
  5. balancing loop: <if can be weakened>
  6. reinforcing loop: <if can be amplified/suppressed>

Proposed intervention: <move the system is considering or you recommend>
System response: <what the system will do in reaction>
Backfire risk: <where intervention could make things worse>

Falsifier: <observable fact that would prove this diagnosis wrong>

output as markdown. include worked example reference: Forrester's Beer Distribution Game & Sterman's 1989 measurement.

Outcome Signal

you know the skill worked when:

• you can name the dominant loop and its classification (R or B) without ambiguity, and explain why observed behavior matches prediction from that loop
• you've located at least one significant delay and can describe why intuition fails across that delay
• you've distinguished stocks from flows in your system and can see why a positive flow might leave a stock dangerously low
• your proposed intervention targets structure or higher on the Meadows hierarchy, not just parameters
• you can articulate what the system will do in response to your intervention, and you've identified a backfire risk
• you have a concrete falsifier: an observable fact that would prove your loop diagnosis wrong (used to validate real outcomes after intervention)

if you're still treating the system as linear, extrapolating recent trends, or proposing parameter-only fixes, the skill has not succeeded.

Loop Patterns Pack

r growth: network effects, viral k>1, compounding learning curves. risk: hits a balancing limit you don't control. defense: identify the limit before you hit it (saturation, regulation, cost structure).

r collapse: death spirals, bank runs, adverse selection cascades. defense: structural circuit breakers, fast intervention, restore balancing signal.

b working: market price discovery, wages, thermostats. don't suppress healthy balancing loops. exhortation ("be more careful") is noise in well-functioning balancing loops.

b + long delay → oscillation: bullwhip in supply chains, cobweb cycles in agriculture, capacity build-out in data centers. defense: shorten delays, damp response, share end-demand data upstream.

r + b → s-curve: technology adoption. see s-curve-technology-adoption. to extend growth, kick off a second reinforcing loop before the first saturates.

Common Rationalizations

Red Flags

• trends extrapolated in a feedback-driven system without regime-switch analysis
• oscillation blamed on external variability without checking internal loop generators
• intervention at parameter level when loop structure is the source
• "be more careful" proposed in a forrester-adversarial structure
• stocks and flows confused in diagnosis
• death spiral or growth narrative with no loop/nodes/delays specified
• all interventions at lowest (parameter) leverage level
• balancing loops suppressed (e.g., ignoring feedback signals to hit targets)
• no falsifier named; diagnosis unfalsifiable

Verification Checklist

• system and variable named explicitly
• at least one reinforcing and one balancing loop identified with full causal chain
• each loop classified by sign-counting; sign-count shown
• delays identified with rough magnitudes
• dominant loop identified; observed behavior consistent with it and its delays
• stocks and flows distinguished
• predicted behavior (without intervention) matches actual; if not, loops re-classified or missed loop identified
• leverage points ranked; recommended intervention not at lowest level if higher leverage is accessible
• system response to intervention considered; backfire risk named
• observable falsifier named (used to validate after intervention)

primary sources: forrester (1961), sterman (1989), meadows (1999).

credits: original framework by clawhub. enriched for implexa standards.