Systems thinking

2026-09-26

How to see and change the structures that produce behavior over time: stocks and flows, feedback loops, delays, causal loop diagrams, the recurring archetypes, Donella Meadows' twelve leverage points, the iceberg model, and the laws that describe how systems resist and surprise their designers. Includes worked math, a step-by-step analysis template and applications to startups, organizations, habits and software. First principles thinking takes a problem apart; this sheet is about how the parts behave once connected. For queueing and failure handling in real services see system design.

What a system is

Donella Meadows (Thinking in Systems, 2008): "A system is an interconnected set of elements that is coherently organized in a way that achieves something." Every system has three kinds of thing:

Part	Meaning	Startup example	How easy to see	Effect of changing it
Elements	the things: people, machines, money, stock, users	engineers, servers, customers, cash	easiest	usually small, unless an element is also a rule-maker
Interconnections	the relationships and information flows that link elements: physical flows, rules, signals	pricing, referral mechanics, pay plans, who reports to whom, dashboards	harder	often large
Purpose or function	what the system actually does, deduced from behavior rather than stated goals	"grow ARR" on the slide; "hit this quarter's number at any cost" in practice	hardest	the largest

Meadows' point: the elements are what we notice, but you can replace every player on a football team and it is still the same team; change the rules or the purpose and it becomes a different system. Judge purpose by what the system does, not by what it says (a company that says "quality first" but promotes whoever ships fastest has a speed purpose).

Term	Meaning
Boundary	what you choose to include; drawn for the question, not given by nature
Environment	everything outside the boundary that affects the system but is not affected by it (in your model)
Open vs closed	whether the system exchanges matter, energy or information with its environment (organizations always do)
State	the values of all the stocks at a moment
Behavior	how the state changes over time; the thing you actually care about
Structure	stocks, flows, feedback loops, delays and rules; structure produces behavior

The core claim of systems thinking: structure produces behavior. Similar structures produce similar behavior whatever the content, so a small vocabulary (stocks, flows, loops, delays) explains a large range of problems.

Stocks and flows

A stock is an accumulation you could count at an instant: water in a bath, money in an account, users, inventory, technical debt, trust, skill. A flow is a rate that changes a stock: liters per minute, pounds per month, sign-ups per week. Stocks change only through flows.

Stock and flows: the bathtub model of a user base with churn

S_{t+1} = S_t + \text{inflow}_t - \text{outflow}_t \qquad\text{or, continuously,}\qquad \frac{dS}{dt} = \text{inflow}(t) - \text{outflow}(t)

Rule	Consequence
a stock rises when inflow exceeds outflow, whatever their absolute sizes	you can grow users by cutting churn as well as by adding sign-ups
dynamic equilibrium: inflow = outflow, so the stock is constant while material moves through it	a "stable" headcount can hide 30% annual turnover
stocks change slowly relative to flows; they are the system's memory and inertia	reputation, skills, culture and debt take years to build or pay down
stocks act as buffers that decouple inflow from outflow	inventory, cash reserves and queues let the two sides run at different rates
you cannot see a stock's future from its level alone; look at the flows	"we have 18 months of runway" depends on the burn rate trend
the time a unit spends in a stock is stock ÷ throughput (Little's law, below)	a 600-ticket backlog handled at 100 a week is a 6-week wait

Worked example: bank balance

Stock: balance, starting at $2,000. Inflow: salary $4,000 a month. Outflow: spending $3,600 a month. Net flow +$400, so the balance grows linearly: $2,400, $2,800, $3,200, … Add interest at 0.4% a month and a reinforcing loop appears (the interest inflow depends on the stock), so growth becomes slightly exponential. Doubling time for a stock growing at $g$ % per period:

t_{\text{double}} \approx \frac{70}{g} \quad\text{(rule of 70; e.g. 7\% a year doubles in about 10 years)}

Worked example: user base with churn

New users arrive at a constant $A$ per month; a constant fraction $c$ of existing users leave each month.

U_{t+1} = U_t + A - cU_t \quad\Rightarrow\quad U^* = \frac{A}{c}, \qquad U_t = U^* + (U_0 - U^*)(1-c)^t

Scenario	Sign-ups $A$	Monthly churn $c$	Steady state $U^*$	Users after 12 months from 0	Half-way time
base	1,000	5%	20,000	9,190	~13.5 months
double acquisition	2,000	5%	40,000	18,390	~13.5 months
halve churn	1,000	2.5%	40,000	10,480	~27 months

Lessons: with constant acquisition, growth always levels off at $A/c$ (a balancing loop through churn). Halving churn has the same effect on the ceiling as doubling acquisition, but takes longer to show, which is why churn work is chronically undervalued in monthly reviews. The gap to steady state closes by a fraction $c$ each month; the half-way time is $\ln 0.5 / \ln(1-c)$ , roughly $0.69/c$ .

Feedback loops

A feedback loop exists when a change in a stock affects the flows into or out of that same stock, directly or through a chain of causes. There are two kinds.

A reinforcing (R) and a balancing (B) loop with link polarities and a delay

	Reinforcing (R, "positive") loop	Balancing (B, "negative") loop
Effect	amplifies change in whichever direction it starts	counteracts change; pushes the stock toward a goal
Behavior alone	exponential growth or collapse	goal seeking; with delays, oscillation
Signature	"the more X, the more X"; virtuous or vicious circle	a gap between actual and desired state drives corrective action
Negative links in the loop	even number (0, 2, …)	odd number (1, 3, …)
Examples	compound interest, network effects, word of mouth, bank runs, tech debt slowing fixes, burnout spirals	thermostat, inventory reordering, autoscaling, hunger, market price adjustment, hiring to a headcount plan
Management trap	assuming growth continues; ignoring the vicious version	fighting a balancing loop you did not see (policy resistance)

"Positive" and "negative" mean the loop's polarity, not good and bad. A reinforcing loop of rising tech debt is positive feedback and very bad news.

Loop dominance: real systems contain many loops, and which one dominates changes over time. S-shaped growth is a reinforcing loop dominating early and a balancing loop (a limit) dominating later. Most surprises are shifts in dominance that nobody modeled.

Delays, non-linearity and bounded rationality

Delays and oscillation

A delay is a lag between cause and effect: between an action and its result, or between a result and your perceiving it. Balancing loops with delays overshoot and oscillate, because the corrective action continues after the gap has closed.

Example	Delay	What happens
Shower	a few seconds between turning the tap and feeling the temperature change	you turn it hotter, feel nothing, turn it more, get scalded, overcorrect to cold, and so on
Beer game (MIT, developed by Jay Forrester's system dynamics group from the 1960s)	order and shipping delays of several weeks at each of four stages: retailer, wholesaler, distributor, factory	a single step increase in customer demand produces large swings in orders and inventory upstream (the bullwhip effect); Sterman (1989) found players typically blame the swings on outside demand rather than their own ordering
Hiring	3–6 months from opening a role to a productive hire	teams hire for last quarter's workload, then have too many people when demand has already turned
Autoscaling	minutes for new instances to boot and warm up	naive scaling on current load overshoots, then scales in too far, then thrashes
Interest rates	monetary policy acts with long and variable lags	central banks that react only to current inflation risk overshooting in both directions

To reduce oscillation	Why it works
act on the supply line (what is already ordered, hired or booting), not just the current gap	stops you ordering the same correction twice
make smaller corrections and wait longer between them	reduces the gain of the balancing loop relative to the delay
shorten the delay (faster information, faster delivery)	the loop sees its own effects sooner
hold buffers (inventory, cash, spare capacity)	absorbs the gap while the correction arrives

Meadows lists delay length, relative to the rate of system change, as leverage point 9: often you cannot shorten the delay, and then slowing the system down so the delay matters less is the fix.

Non-linearity

Effects are rarely proportional to causes. Relationships have thresholds, saturation and tipping points: the first server doubling capacity cuts latency a lot, the fifth barely at all; a queue at 50% utilization is fine, at 95% it explodes (see the software section). Linear extrapolation from the current range is the most common modeling error.

Bounded rationality

Herbert Simon's term: people make reasonable decisions given the limited, delayed and local information they have. In systems terms, each actor can behave rationally and the system still produce an outcome nobody wants (fishers overfishing, teams hoarding shared staff, suppliers over-ordering in the beer game). Meadows' conclusion is that replacing the people rarely helps; changing the information and incentives they see does.

Causal loop diagrams

A causal loop diagram (CLD) shows variables joined by arrows labeled with polarity, and the loops those arrows form. It is a thinking tool: cheap, qualitative and good for conversations. For numbers you need a stock-and-flow model (spreadsheet, or tools such as Vensim, Stella or Insight Maker).

Notation	Meaning
variable	a noun phrase that can go up or down: "customer satisfaction", not "satisfy customers"
arrow A → B	a change in A causes a change in B, all else equal
+ (or s, "same")	A up → B up, and A down → B down, compared with what B would otherwise have been
− (or o, "opposite")	A up → B down, and A down → B up
‖ across an arrow	a significant delay
R / B in a loop	reinforcing / balancing loop, often with a name ("R1: word of mouth")

Polarity rule: count the negative links around a loop. Even (including zero) → reinforcing. Odd → balancing. Check by walking the loop: assume the first variable rises, follow the signs around, and see whether you come back with "rises" (reinforcing) or "falls" (balancing).

Start from a behavior over time
Pick one or two variables whose trend is the problem and sketch the graph: "support tickets per week, rising for 9 months". The diagram must explain that shape.
Name the key variables
Write 5–10 nouns that can increase or decrease. Use the positive sense ("morale", not "low morale"). Include goals, perceived states and pressures as well as physical quantities.
Draw causal links and label polarity
For each link ask "if A goes up, does B go up or down, all else equal?" Only draw direct causes. Mark significant delays.
Close the loops and label them
Find the circles. Count negative links to label each R or B. Name each loop in a few words.
Tell the story and check it against the data
Walk each loop aloud. Does the dominant loop explain the observed trend? What would have to change for another loop to dominate? If the diagram cannot reproduce the behavior, it is missing a loop, a delay or a limit.
Look for leverage
Mark where you could add, cut or weaken a link, change a goal, shorten a delay or change who sees what information. Check each intervention for side effects by walking the loops again.

Common CLD error	Fix
variables that are actions or events ("launch campaign")	make them quantities ("marketing spend")
ambiguous polarity ("it depends")	the link hides two mechanisms; split it into two variables
everything connected to everything	keep 5–15 variables per diagram; one diagram per question
correlation drawn as causation	draw a link only if you can describe the mechanism
no delays marked	ask of every link "how long before the effect shows?"
no goals shown	balancing loops need a goal or desired state; draw it explicitly

Behavior over time patterns

A small number of structures produce almost all observed behavior (Sterman, Business Dynamics, 2000, calls these the fundamental modes of dynamic behavior). Match the graph to the structure, then look for that structure.

Pattern	Shape	Structure that produces it	Simple model	Example
Exponential growth or decay	curving ever steeper up (or down)	a dominant reinforcing loop	$S_{t+1} = S_t(1+g)$	compound interest, early viral growth, a bank run
Goal seeking	rises or falls quickly, then flattens at a target	a dominant balancing loop	$S_{t+1} = S_t + k(G - S_t)$	coffee cooling to room temperature, filling a hiring plan
Oscillation	repeated overshoot above and below a target	a balancing loop with a significant delay	goal seeking where the correction acts on an old gap	shower temperature, inventory cycles, commodity price cycles
S-shaped growth	exponential start, then levelling off	a reinforcing loop meeting a balancing loop (a limit)	logistic: $\dfrac{dS}{dt} = rS\left(1 - \dfrac{S}{K}\right)$	product adoption in a finite market, a new team's output
S-shaped growth with overshoot	grows past the limit, then oscillates around it	S-shaped growth where the limiting loop is delayed	logistic with a delay	a population rising past its food supply, hiring past demand
Overshoot and collapse	grows past the limit and falls sharply	growth that erodes or consumes its own limit (a resource, trust, a customer base)	the carrying capacity $K$ itself falls when exceeded	fisheries, a platform that burns its users' goodwill with ads, a credit bubble

System archetypes

Archetypes are recurring combinations of loops with a known story and known leverage. Peter Senge popularised them in The Fifth Discipline (1990), building on work at Innovation Associates and MIT; Daniel Kim and the journal The Systems Thinker developed the practitioner versions, and William Braun's 2002 summary lists ten. Accidental Adversaries comes from Jennifer Kemeny (The Systems Thinker, 1994).

Archetype	Structure	Symptom (what you hear)	Example	Leverage
Limits to Growth	R (growth engine) runs into B (a limiting condition)	"we're working harder but growth has stalled"	a product grows through referrals until the reachable market or support capacity saturates	find and remove or weaken the limit before it binds; stop pushing the growth engine harder
Shifting the Burden	a quick symptomatic fix (B1) and a slow fundamental fix (B2); the quick fix has a side effect that weakens the fundamental one	"it keeps coming back, but the quick fix works for now"	senior engineers firefight every incident, so the team never learns to; discounts instead of a better product	use the symptomatic fix only to buy time; invest in the fundamental solution; watch for dependency
Eroding Goals (drift to low performance)	gap between goal and reality closed by lowering the goal rather than by corrective action	"that target was unrealistic anyway"	on-call SLOs quietly relaxed each quarter; "good enough" test coverage sliding	anchor goals to external benchmarks or customers; make the history of the goal visible
Escalation	two balancing loops, each party responding to its position relative to the other, forming a reinforcing figure of eight	"we have to respond to what they just did"	price wars, feature arms races, hospital building races, arms races	find a way for one side to de-escalate unilaterally, or reframe the goal so it is not relative
Success to the Successful	two reinforcing loops competing for a shared resource: success brings resources, which bring more success	"they've earned it"; "the other team never delivers"	the flagship product gets all the engineers while a promising new one starves; network effects	question the allocation rule; separate goals and budgets; level the playing field for new entrants
Tragedy of the Commons	each actor's reinforcing gain from using a shared resource; total use depletes it (a shared balancing loop with a delay)	"why is the shared service so slow for everyone?"	a shared platform team, a shared database, CI runners, a fishery, a common pasture	make the cumulative cost visible to each user; charge for or allocate use; govern the commons (Ostrom)
Fixes that Fail	a quick fix relieves the symptom (B) but causes delayed unintended consequences that worsen it (R)	"it worked at first, now it's worse than before"	adding people to a late project (Brooks's law); cutting maintenance to hit a budget; retries that amplify an outage	map side effects before acting; watch for the delay; prefer fixes that address the cause
Growth and Underinvestment	Limits to Growth where the limit is capacity that could be expanded, but demand falls first because of poor performance, which "proves" investment is not needed	"demand is soft, so we can't justify more capacity"	a startup that never hires support, so slow response loses customers, so it never "needs" support	invest in capacity ahead of demand based on external signals; hold performance standards fixed
Accidental Adversaries	two partners whose own fixes unintentionally obstruct each other, turning a collaborative reinforcing loop into mutual damage	"they keep undermining us" (from both sides)	Procter & Gamble's promotions and Walmart's forward-buying, as told by Kemeny	bring both parties together to map the whole system; agree shared measures (P&G moved to everyday low pricing)

Meadows' Thinking in Systems has a parallel list she calls system traps: policy resistance, tragedy of the commons, drift to low performance, escalation, success to the successful, shifting the burden to the intervenor, rule beating, and seeking the wrong goal. Rule beating (following the letter of a rule to defeat its purpose) and seeking the wrong goal (optimizing the measured proxy) are the two most common in companies; they are Goodhart's law in systems language.

Leverage points

From Donella Meadows, "Leverage Points: Places to Intervene in a System" (1999). The list runs from least to most effective. The wording in the first column is Meadows' own summary list; the examples are ours.

#	Leverage point (Meadows' wording)	Company example
12	Constants, parameters, numbers (such as subsidies, taxes, standards).	tweaking a discount from 10% to 15%, a sales quota, an SLO target
11	The sizes of buffers and other stabilizing stocks, relative to their flows.	cash runway, inventory, spare server capacity, slack in the roadmap
10	The structure of material stocks and flows (such as transport networks, population age structures).	the physical architecture: data centers, office locations, the org's seniority mix
9	The lengths of delays, relative to the rate of system change.	deploy lead time, time to hire, how fast sales feedback reaches product
8	The strength of negative feedback loops, relative to the impacts they are trying to correct against.	incident review strength, code review, customer complaint escalation, auditors
7	The gain around driving positive feedback loops.	referral incentives, network effects, compounding reinvestment, how fast tech debt compounds
6	The structure of information flows (who does and does not have access to information).	showing engineers the cloud bill or customer tickets; public dashboards; transparent pay bands
5	The rules of the system (such as incentives, punishments, constraints).	compensation plans, promotion criteria, approval rules, pricing model
4	The power to add, change, evolve, or self-organize system structure.	teams allowed to form, split and change their own process; an internal platform others build on
3	The goals of the system.	"growth at all costs" vs "profitable growth"; "ship features" vs "retain customers"
2	The mindset or paradigm out of which the system — its goals, structure, rules, delays, parameters — arises.	shared beliefs such as "software is a cost center" or "customers are the enemy of margin"
1	The power to transcend paradigms.	holding no model as final truth; being able to switch frames as the situation demands

Observation from the essay	Practical reading
Meadows says parameters are "dead last" on the list, yet get most of our attention (she guesses "90, no 95, no 99 percent")	most management meetings argue about numbers; the structure producing them goes unexamined
she quotes Jay Forrester: people often know intuitively where leverage points are, but push them "IN THE WRONG DIRECTION"	a leverage point found by intuition still needs its direction checked against the loops
the list began as a nine-item list scribbled at a meeting and was revised over years; she calls it "a work in progress"	treat the order as a heuristic, not a law
information flows (6) are cheap and powerful: "Missing feedback is one of the most common causes of system malfunction."	before changing rules, make the consequences visible to the people causing them

The iceberg model

A teaching device used widely in systems thinking courses: what you see (events) is the tip; below it are patterns, the structures producing them, and the mental models that sustain the structures. No single originator is documented; Senge's The Fifth Discipline distinguishes event, pattern-of-behavior and structural explanations, and practitioner versions add mental models as the deepest layer.

Level	Question	Typical response	Example: production outages
Events	what just happened?	react: fix it now	the site went down on Tuesday
Patterns	what has been happening over time?	anticipate: plan for it	outages cluster after big Friday releases; three this quarter
Structures	what is causing the pattern?	design: change the system	weekly batch releases, no canarying, on-call rota staffed by the release authors, incentives to ship by the sprint deadline
Mental models	what beliefs keep the structure in place?	transform: change the thinking	"moving fast means skipping process"; "outages are bad luck"; "ops is someone else's job"

Leverage rises as you go down; so does the difficulty and time needed.

Emergence, resilience, self-organization and hierarchy

Property	Meaning	Implication	Example
Emergence	system-level behavior that none of the parts has alone and that comes from their interactions	you cannot understand it by studying parts in isolation; you can only change it via interactions	traffic jams, market prices, team culture, a flash crash
Resilience	the ability to survive and recover from disturbance, from a variety of redundant balancing loops (Meadows)	it is often invisible until lost; optimizing for efficiency strips it out	just-in-time supply chains in 2020–21; a single engineer who knows the billing system
Self-organization	the capacity to make its own structure more complex: learn, diversify, evolve	produces surprises; suppressing it for control reduces adaptability	open-source ecosystems, internal tooling that spreads team to team
Hierarchy	subsystems nested in larger systems, each mostly self-regulating, with fewer links between than within them	makes complex systems stable and evolvable; fails when the top over-controls (centralization) or the parts optimize against the whole (suboptimisation)	microservices, divisions, cells in organs

Herbert Simon's "The Architecture of Complexity" (1962) makes the hierarchy argument with a parable of two watchmakers: the one who builds from stable sub-assemblies finishes watches; the one who builds each watch as a single assembly loses all progress whenever he is interrupted (paraphrased). The same logic underlies Gall's law, below.

Laws, effects and traps

Law or effect	Statement	Use
Gall's law	John Gall, General Systemantics (1975): "A complex system that works is invariably found to have evolved from a simple system that worked. A complex system designed from scratch never works and cannot be made to work. You have to start over, beginning with a working simple system." (wording varies slightly between editions)	ship the simplest working version and grow it; distrust big-bang rewrites and grand org redesigns
Goodhart's law	Charles Goodhart (1975): "Any observed statistical regularity will tend to collapse once pressure is placed upon it for control purposes." Marilyn Strathern's 1997 generalization: "When a measure becomes a target, it ceases to be a good measure."	pair every target with a counter-metric; rotate metrics; keep some measures for observation only
Campbell's law	Donald Campbell (1976; published 1979): "The more any quantitative social indicator is used for social decision-making, the more subject it will be to corruption pressures and the more apt it will be to distort and corrupt the social processes it is intended to monitor."	the social version of Goodhart: expect gaming, not just drift
Conway's law	Melvin Conway, "How Do Committees Invent?" (Datamation, 1968): "organizations which design systems (in the broad sense used here) are constrained to produce designs which are copies of the communication structures of these organizations."	design team boundaries to match the architecture you want (the "inverse Conway maneuver")
Brooks's law	Fred Brooks, The Mythical Man-Month (1975): "Adding manpower to a late software project makes it later."	a Fixes that Fail instance: onboarding and communication costs arrive before the extra capacity
Second-order effects	the consequences of the consequences; first-order effects are usually intended, later ones often not	for each intervention ask "and then what?" at least twice
Unintended consequences / perverse incentives	an incentive rewards the measured proxy, so people produce the proxy	pay for outcomes that are hard to fake; test incentives on a small scale first
Policy resistance	a system pushes back against an intervention because its actors keep pulling toward their own goals (Meadows' trap; Senge: "compensating feedback")	find out what each actor wants and why; align goals rather than pushing harder

The cobra effect, the story of a colonial bounty on cobras in Delhi leading to cobra farming, is the standard illustration of perverse incentives, but it is anecdotal: the term was coined by the economist Horst Siebert (Der Kobra-Effekt, 2001) and no contemporary record of cobra breeding has been found. The better-documented case is the Great Hanoi Rat Massacre of 1902, researched by historian Michael Vann: the French administration paid a bounty per rat tail, and people cut off tails and released the rats to breed.

Le Chatelier's principle in chemistry says a system at equilibrium responds to a disturbance by shifting to counteract it. Policy resistance is the social analogue, and the reason Jay Forrester called social systems "counterintuitive" ("Counterintuitive Behavior of Social Systems", 1971): intuitive policies often attack symptoms, trigger compensating loops, and leave the problem as bad or worse.

Senge's "laws of the fifth discipline" (The Fifth Discipline, ch. 4), condensed and paraphrased:

Law (paraphrased)	Loop behind it
today's problems come from yesterday's solutions	Fixes that Fail, Shifting the Burden
the harder you push, the harder the system pushes back	policy resistance, compensating balancing loops
behavior grows better before it grows worse	delayed side effects
the easy way out usually leads back in; the cure can be worse than the disease	symptomatic fixes that erode the fundamental solution
faster is slower	pushing past a system's optimal rate triggers limits
cause and effect are not closely related in time and space	delays, and effects that appear in another part of the system
small changes can produce big results, but the highest-leverage areas are often the least obvious	leverage points
dividing an elephant in half does not produce two small elephants	emergence: the whole has properties the parts lack
there is no blame	the structure, not the individual, produces the behavior

Applying it

To a startup

Question	Systems framing	What to look at
Why has growth stalled?	stock of customers near its steady state $A/c$ , or a growth loop meeting a limit	compute $A/c$ ; list the limits (market size, support capacity, onboarding friction, channel saturation)
Which growth loops do we have?	reinforcing loops: referral (users → invites → users), content (users → content → search traffic → users), paid (revenue → ad spend → customers → revenue)	the gain of each loop (e.g. invites per user × conversion) and its delay; a loop with gain below 1 per cycle does not self-sustain
Why does churn keep rising as we grow?	Growth and Underinvestment: support, onboarding and reliability capacity lag demand	response times, incident rate and churn by cohort; invest ahead of the limit
Why do discounts not stick?	Shifting the Burden: discounting relieves the sales gap but trains customers to wait for discounts and reduces pressure to improve the product	share of revenue on discount over time; win-rate at full price
Why is our metric improving but the business not?	Goodhart and seeking the wrong goal	a counter-metric for every target (activation with 30-day retention, velocity with change failure rate)

To an organization

Symptom	Likely structure	Intervention
a shared platform team is overwhelmed and every team complains	Tragedy of the Commons	visible cost per requesting team, a published intake policy, self-serve tooling
senior people are always firefighting	Shifting the Burden to the intervenor	pair on incidents, runbooks, rotate on-call through the whole team
targets quietly lowered each quarter	Eroding Goals	anchor to customer-facing or external benchmarks; publish the target's history
teams "throwing work over the wall"	Conway's law plus Accidental Adversaries	redraw team boundaries around the product flow; shared goals across the handoff
reorganizations every year with no lasting change	intervening at the level of parameters and elements, not goals, rules or information flows	change what information people see and what they are rewarded for

To personal habits

Habit dynamic	Structure	Lever
fitness, skill and savings build slowly and decay slowly	stocks with small flows and long delays	judge the inflow (sessions, hours, deposits), not the stock, week to week
a streak keeps itself going	reinforcing loop: practice → competence → enjoyment → practice	make the first cycles easy so the loop starts
caffeine to cover lost sleep	Shifting the Burden: the quick fix worsens the underlying sleep deficit	fix the fundamental (sleep time) and use the quick fix only while transitioning
"I'll run 3 times a week" becomes once, then "when I can"	Eroding Goals	fix the goal to an external anchor (a race date, a training partner)
weight regain after a diet	balancing loops (appetite, metabolic adaptation) resisting a parameter change	change the structure (environment, routines), not just the target number

To software

Queues, retries and autoscalers are feedback systems; most large outages are loops nobody drew. See system design for the architecture patterns.

Concept	Systems view	Numbers or rule
Little's law	stock = flow × time in the stock	$L = \lambda W$ : 200 req/s × 0.05 s = 10 requests in flight
Queueing and utilization	non-linear: waiting time explodes as utilization $\rho$ approaches 1	single-server M/M/1 queue: $W = S/(1-\rho)$ ; with a 10 ms service time, 20 ms at 50%, 100 ms at 90%, 1 s at 99%
Backpressure	a balancing loop: a full downstream buffer slows the upstream producer	bounded queues, rejecting or slowing input when full, rather than unbounded buffering
Retry storms	a reinforcing loop: failures → retries → more load → more failures	3 retries per call means up to 4× load on a failing service; 3 retrying layers compound to up to $4^3 = 64$ ×
Exponential backoff with jitter	weakens the retry loop's gain and spreads its timing	cap attempts; randomize delays so clients do not retry in sync
Circuit breakers and load shedding	add a balancing loop that cuts load when error rates cross a threshold	fail fast, serve degraded responses, protect the constrained resource
Autoscaling with warm-up delay	balancing loop with a delay: oscillates if tuned aggressively	scale on leading signals, add cool-down periods, keep a buffer of warm capacity
Metastable failures	a system that stays broken after the trigger is removed because a reinforcing loop (retries, cache misses, queue growth) sustains the overload (Bronson et al., HotOS 2021)	the fix is to break the loop (shed load, drop queues, disable retries), not to wait
Caches	a buffer stock; its failure removes a balancing loop protecting the database	cold-cache start after an outage can overload the origin: warm caches gradually

Systems analysis template

SYSTEMS ANALYSIS: <problem>
 
1. BEHAVIOR OVER TIME
   Variable(s) that show the problem, with units:
   Sketch the trend (past 1-3 years) and the desired trend:
   Pattern: growth / decline / goal-seeking / oscillation /
            S-curve / overshoot-and-collapse
 
2. BOUNDARY AND HORIZON
   What is inside the system? What is environment?
   Time horizon long enough to see the delays:
 
3. STOCKS AND FLOWS
   Stock            Inflows              Outflows
   ..............   ..................   .................
   Implied steady state (inflow / outflow rate):
 
4. FEEDBACK LOOPS (causal loop diagram)
   R1 ................  (links, polarity, delays)
   B1 ................  (goal it seeks, delay)
   Which loop dominates now? Which will dominate later?
 
5. ARCHETYPE MATCH (if any)
   Limits to Growth / Shifting the Burden / Eroding Goals /
   Escalation / Success to the Successful / Tragedy of the
   Commons / Fixes that Fail / Growth and Underinvestment /
   Accidental Adversaries
 
6. ICEBERG
   Events:        ...
   Patterns:      ...
   Structures:    ...
   Mental models: ...
 
7. ACTORS AND BOUNDED RATIONALITY
   Actor   Goal   Information they see   Incentive
   Why is each actor's behavior locally rational?
 
8. LEVERAGE POINTS (Meadows 12 -> 1)
   Candidate interventions, with the level of each:
   Parameter / buffer / delay / loop strength / information /
   rules / self-organization / goals / paradigm
 
9. SIDE EFFECTS AND SECOND-ORDER EFFECTS
   For each intervention: "and then what?" twice.
   Which loop might push back (policy resistance)?
   What gets gamed if this becomes a target (Goodhart)?
 
10. TEST AND MONITOR
    Smallest reversible experiment:
    Leading indicator + counter-metric:
    Expected delay before the effect shows:
    Review date:

Common mistakes

Mistake	Why it hurts	Instead
Drawing everything	a 60-variable diagram explains nothing and persuades no one	one question, one diagram, 5–15 variables; the boundary is set by the question
Ignoring delays	you misread slow results as failure, double the dose, and overshoot	mark every significant delay; set the review date after the delay, not before
Treating symptoms	the fix relieves pressure and the cause persists or grows	ask which loop produces the symptom; check for Shifting the Burden and Fixes that Fail
Blaming people	replacing people inside the same structure reproduces the behavior	ask why the behavior is locally rational; change information, rules or goals
Linear extrapolation	the dominant loop will change; limits appear	look for the limit and the balancing loop that will take over
Confusing stocks and flows	"revenue fell" (a flow) treated like "customers fell" (a stock)	label units: a stock is "X"; a flow is "X per period"
Only intervening at parameters	the cheapest intervention is the weakest	walk down the leverage list before settling on a number change
Systems thinking as an excuse for inaction	"it's complex" becomes a reason not to act	pick the smallest reversible intervention and learn from it
Unfalsifiable diagrams	a CLD can explain any outcome after the fact	write down in advance what the model predicts; check it

Critiques and limits

Critique	Substance	Response
Qualitative diagrams are not models	CLDs cannot tell you magnitudes, timing or which loop wins; two people can draw opposite diagrams	quantify the key stocks and flows in a spreadsheet or a system dynamics tool before big decisions
Easy to be vague	"everything is connected" is true and useless	insist on units, polarities, delays and a testable prediction
Hindsight fitting	archetypes can be matched to almost any story after the fact	use them prospectively: predict what happens next, then check
Boundary choice is subjective	what you leave out drives the conclusion	state the boundary; test whether widening it changes the answer
The Limits to Growth controversy	the 1972 world model that launched much of the field was heavily criticized by economists for omitting prices, substitution and technical change	models are aids to thinking about structure; they are not forecasts
Tragedy of the Commons is not inevitable	Elinor Ostrom (Governing the Commons, 1990) documented many communities that manage shared resources sustainably through local rules, monitoring and sanctions	the archetype describes an unmanaged commons; governance is a leverage point
Can ignore power and politics	diagrams of "the system" can hide who benefits from the current structure	include actors and their goals explicitly (template step 7)
Over-complication	for a simple, linear, well-understood problem, a loop diagram is overhead	use it when behavior is recurring, counterintuitive or resists fixes

References

Donella H. Meadows, Thinking in Systems: A Primer, ed. Diana Wright (Chelsea Green, 2008): definition of a system, stocks and flows, loops, resilience, self-organization, hierarchy, system traps
Donella Meadows, "Leverage Points: Places to Intervene in a System" (1999), Donella Meadows Project (opens in a new tab): the twelve leverage points and the original nine-point list
The Donella Meadows Project (Academy for Systems Change) (opens in a new tab): Meadows' essays, columns and systems-thinking resources
Peter M. Senge, The Fifth Discipline: The Art and Practice of the Learning Organization (Doubleday/Currency, 1990): archetypes (appendix), the laws of the fifth discipline, levels of explanation
Peter M. Senge et al., The Fifth Discipline Fieldbook (Doubleday/Currency, 1994): practitioner archetype material, including Accidental Adversaries
William Braun, "The System Archetypes" (2002) (opens in a new tab): ten archetypes with generic diagrams, behavior over time and prescriptive actions, drawing on Daniel Kim's Systems Thinker articles
Jennifer Kemeny, "'Accidental Adversaries': When Friends Become Foes", The Systems Thinker (opens in a new tab): the archetype and the P&G–Walmart example
The Systems Thinker (archive) (opens in a new tab): Pegasus Communications' practitioner journal on archetypes and causal loop diagrams
Daniel H. Kim and Virginia Anderson, Systems Archetype Basics (Pegasus Communications, 1998): workbook on the archetypes
Jay W. Forrester, Industrial Dynamics (MIT Press, 1961): the origin of system dynamics
Jay W. Forrester, "Counterintuitive Behavior of Social Systems", Theory and Decision 2 (1971): why intuitive policies fail in feedback systems
John D. Sterman, Business Dynamics: Systems Thinking and Modeling for a Complex World (Irwin/McGraw-Hill, 2000): the standard textbook; modes of behavior, stock-and-flow modeling
John D. Sterman, "Modeling Managerial Behavior: Misperceptions of Feedback in a Dynamic Decision Making Experiment", Management Science (1989): the beer game findings
Wikipedia: Beer distribution game (opens in a new tab): rules and history
Wikipedia: Causal loop diagram (opens in a new tab): notation and the polarity rule
Herbert A. Simon, "The Architecture of Complexity", Proceedings of the American Philosophical Society (1962): hierarchy and the watchmakers parable
John Gall, General Systemantics (1975; later Systemantics and The Systems Bible): Gall's law
Wikiquote: John Gall (opens in a new tab): the Gall's law passage and its edition variants
Wikipedia: Goodhart's law (opens in a new tab): Goodhart's 1975 wording, Hoskin (1996) and Strathern (1997)
Marilyn Strathern, "'Improving ratings': audit in the British University system", European Review (1997) (opens in a new tab): source of "When a measure becomes a target…"
Donald T. Campbell, "Assessing the Impact of Planned Social Change", Evaluation and Program Planning (1979) (opens in a new tab): Campbell's law
Melvin E. Conway, "How Do Committees Invent?", Datamation (1968) (opens in a new tab): Conway's law
Frederick P. Brooks Jr., The Mythical Man-Month (Addison-Wesley, 1975): Brooks's law
Wikipedia: Perverse incentive (cobra effect) (opens in a new tab): Siebert's coinage, the lack of evidence for the Delhi story, and the Hanoi rat bounty
Michael G. Vann, "Of Rats, Rice, and Race: The Great Hanoi Rat Massacre", French Colonial History (2003): the documented bounty case
Garrett Hardin, "The Tragedy of the Commons", Science (1968), and Elinor Ostrom, Governing the Commons (Cambridge University Press, 1990): the commons and its critique
Nathan Bronson et al., "Metastable Failures in Distributed Systems", HotOS (2021): self-sustaining overload loops in software