MIT OpenCourseWare · Published 2021-11-17

System Dynamics: Systems Thinking and Modeling for a Complex World

privateReview: unreviewedConfidence: high
system dynamicssystems thinkingFeedback loopsStocks and flowsMental modelscontrol theoryDelayspolicy resistanceagent architecture
Open on YouTube ↗

Summary

Overview

  • Speaker: James Edward Paine
  • Channel: MIT OpenCourseWare
  • Main topic: System Dynamics and Systems Thinking
  • Purpose: Provide a comprehensive overview of system dynamics, systems thinking, tools, and modeling frameworks for understanding complex social and economic systems. An introductory workshop on System Dynamics and Systems Thinking, covering foundational concepts, tools such as causal loop diagrams and stock and flows, open loop versus closed loop thinking, the role of mental models, and hands-on simulation overviews.

Topic Map

Introduction and Agenda

  • Explanation: Introduction of the instructor, James Edward Paine, and an overview of the workshop structure including systems thinking, simulations, and debriefs.
  • Key claims:
    • System dynamics is control theory applied to social systems.
    • There is a subtle difference between system dynamics as a method set and systems thinking as a field.
  • Examples:
    • Management flight simulator usage in workshop
  • Terminology:
    • System Dynamics
    • Systems Thinking
    • Behavioral Operations Management
  • Why it matters: Sets expectations for the workshop and introduces foundational definitions.

History of System Dynamics

  • Explanation: Overview of Jay Forrester founding the field at MIT in the 1950s, bridging computer science, control theory, and social systems.
  • Key claims:
    • Jay Forrester created system dynamics based on his background in electrical engineering and digital computers.
    • Social systems are much harder to control than physical systems.
  • Examples:
    • Whirlwind I computer
    • World2 simulation for Club of Rome
    • Limits to Growth
  • Terminology:
    • Jay Forrester
    • Industrial Dynamics
    • Control Theory
    • Limits to Growth
  • Why it matters: Provides historical context and origins of system dynamics in engineering and computing.

Open Loop vs. Systems Thinking

  • Explanation: Contrast between traditional linear project management (identify problem, gather data, evaluate, select solution, implement) and closed-loop feedback systems thinking.
  • Key claims:
    • No decision you make exists in isolation.
    • There are no side effects; there are simply effects you have not thought about yet.
  • Examples:
    • Cutting someone off on the highway and attributing it to bad drivers rather than systemic/situational causes.
  • Terminology:
    • Open Loop Thinking
    • Closed Loop
    • Feedback Loop
    • Side Effects
    • Fundamental Attribution Error
  • Why it matters: Highlights common cognitive biases and the need to perceive feedback loops in complex environments.

Foundations of Systems Thinking

  • Explanation: Core principles including structure generating behavior, mental models mattering, and avoiding the fundamental attribution error.
  • Key claims:
    • Structure generates behavior.
    • Dynamics emerge from the interaction of physics, information availability, and decision rules.
    • It's not enough to change physical structure; mental models and incentives must be addressed.
  • Examples:
    • Oil price boom and bust cycles governed by underlying market structures rather than isolated events.
  • Terminology:
    • Structure Generates Behavior
    • Mental Models
    • Fundamental Attribution Error
    • Visibility
  • Why it matters: Explains why people often blame individuals rather than the systemic structures causing undesirable behaviors.

Tools and Methods: Causal Links and Loops

  • Explanation: Introduction to causal links, reinforcing loops, and balancing loops.
  • Key claims:
    • Causal links define how one variable changes another.
    • Reinforcing loops amplify change; balancing loops seek goals or balance.
    • Counting negatives in a loop determines loop polarity (odd = balancing, even = reinforcing).
  • Examples:
    • Production, inventory, and shipments causal links
    • Employee skill, customer satisfaction, and complaints reinforcing loop
  • Terminology:
    • Causal Link
    • Reinforcing Loop
    • Balancing Loop
    • Goal Seeking Loop
    • Loop Polarity
  • Why it matters: Provides foundational diagrammatic tools for mapping feedback structures.

Tools and Methods: Stock and Flows

  • Explanation: Explanation of stocks (accumulations with memory) and flows (inflows and outflows) using the bathtub/hydraulic metaphor.
  • Key claims:
    • A stock is anything that accumulates over time and has memory.
    • To change a stock, you must alter inflows or outflows.
    • Interest rate is a price (stock-related), not a flow, despite its name.
  • Examples:
    • Greenhouse gases in the atmosphere
    • Balance sheet vs. cash flow statement
  • Terminology:
    • Stock
    • Flow
    • Hydraulic Metaphor
    • Integral Representation
    • Differential Representation
  • Why it matters: Stocks and flows form the backbone of quantitative system dynamics modeling.

Software and Resources

  • Explanation: Overview of software tools like Vensim and Stella Architect, along with textbooks, websites, and courses at MIT.
  • Key claims:
    • Vensim and Stella Architect are standard software packages for system dynamics modeling.
    • All models are wrong, but some models are useful.
  • Examples:
    • Business Dynamics by John Sterman
    • En-Roads climate policy simulator
    • Creative Learning Exchange
  • Terminology:
    • Vensim
    • Stella Architect
    • NetLogo
    • Business Dynamics
    • En-Roads
  • Why it matters: Guides learners to resources and software for further study and modeling practice.

Key Points

Structure Generates Behavior

  • Explanation: The observable events and patterns in a system are driven by underlying physical structures, information availability, and decision rules.
  • Evidence: Oil price volatility analysis showing chronic boom and bust cycles across decades.
  • Practical implication: When trying to fix a persistent problem, modifying the underlying structure and policies is more effective than blaming individuals.

Mental Models Matter

  • Explanation: Decision-makers operate based on mental models that filter information and shape actions.
  • Evidence: Discussions on behavioral operations management and supply chain bullwhip effects.
  • Practical implication: Effective interventions require eliciting, articulating, and expanding mental models to account for feedback and delays.

The Importance of Feedback and Delays

  • Explanation: Systems feature tight coupling and time delays between decisions and system state changes, leading to unintended consequences.
  • Evidence: Open loop vs closed loop diagrams showing how decisions alter system states which feed back into future decisions.
  • Practical implication: Failing to account for delays leads to overshooting and policy resistance.

Frameworks, Models & Processes

Spiral Approach to Model Formulation

  • How it works: An iterative process of problem articulation, dynamic hypothesis, formulation, testing, and policy evaluation.
  • Components:
    • Problem Articulation (Boundary Selection)
    • Dynamic Hypothesis
    • Formulation
    • Testing
    • Policy Formulation & Evaluation
  • When to use: When developing system dynamics models and tackling complex organizational or policy problems.

Causal Loop Diagrams (CLDs)

  • How it works: Mapping variables connected by directed arrows labeled with positive or negative signs to indicate causal relationships.
  • Components:
    • Variables
    • Causal Links (+/-)
    • Reinforcing Loops (R)
    • Balancing Loops (B)
  • When to use: To capture mental models, feedback structures, and qualitative causal relationships.

Stock and Flow Diagrams (Compartmental Models)

  • How it works: Quantifying system states as stocks connected by flows governed by rates, valves, and sources/sinks.
  • Components:
    • Stocks (Levels)
    • Flows (Rates/Valves)
    • Cloud symbols (boundaries)
    • Converters/Auxiliaries
  • When to use: When building quantitative simulation models that require mathematical tracking of accumulations over time.

Examples & Case Studies

A trader causing spikes in oil prices followed by prolonged drops and OPEC rumors.

  • Illustrates: Events vs. patterns of behavior in commodity markets.
  • Lesson: Looking only at isolated events misses the underlying cyclical structure of commodity markets.

The Vasa warship tipping over on its maiden voyage due to late design modifications adding too much top-weight.

  • Illustrates: Unintended consequences and compounding design changes in complex projects.
  • Lesson: Late-stage changes driven by shifting goals can destabilize an otherwise sound structural system.

Actionable Takeaways

  • Immediate:
    • Distinguish between open-loop linear thinking and closed-loop feedback thinking.
    • Identify reinforcing and balancing loops in everyday business problems.
    • Recognize stocks (accumulations with memory) versus flows.
  • Strategic:
    • Focus on shifting structural policies rather than blaming individuals for systemic outcomes.
    • Use simulation to test mental models and evaluate high-leverage policy choices.
    • Embrace an iterative spiral approach to modeling complex systems.
  • Questions to investigate:
    • What are the feedback loops operating in my organization?
    • Where are the critical delays between decisions and outcomes?
    • How do current mental models constrain effective problem-solving?

Claims Worth Verifying

  • Jay Forrester created system dynamics in 1958 with the publication of Industrial Dynamics. (historical fact)
  • Jay Forrester invented magnetic core memory (RAM) and worked on Whirlwind I. (historical fact)

Notable Quotes

"System dynamics: systems thinking and modeling for a complex world." (at 0:27) "Structure generates behavior." (at 30:21) "All models are wrong, but some models are useful!" (at 41:45)

Compressed Summary

  • System dynamics applies control theory and feedback loops to social and economic systems.
  • Structure generates behavior; mental models and feedback loops dictate system outcomes.
  • Tools include causal loop diagrams and stock-and-flow compartmental models.
  • Modeling is an iterative spiral process focused on policy design rather than exact point prediction.
  • Keywords: system dynamics, systems thinking, feedback loops, stocks and flows, mental models
  • Core insight: Understanding complex systems requires shifting from linear open-loop thinking to analyzing closed-loop feedback structures that generate behavior over time.

Core insights

5
Mental Modelmedium noveltystrong evidence

Persistent system behavior is generated by underlying structure (physical paths, information availability, decision rules), not by isolated events or individual actors. Interventions should therefore target structure and policy rather than blame components, which is a direct analog to debugging agent systems.

Why it matters

For agent engineers, repeated failures of a component are often symptoms of the surrounding architecture: context visibility, reward/incentive rules, tool feedback, or decision rules. Fixing the structure is more durable than blaming the model or a single node.

Generalization

When a complex system reliably produces undesirable behavior, redesign the feedback and decision structure instead of swapping individual components.

Structure generates behavior.
Open source video
Dynamics emerge from the interaction of physics, information availability, and decision rules.
Open source video
When trying to fix a persistent problem, modifying the underlying structure and policies is more effective than blaming individuals.
Open source video
Mental Modelhigh noveltystrong evidence

Reframing 'side effects' as 'effects you have not thought about yet' forces engineers to treat every agent action as part of a closed loop where the action changes system state and that state feeds back into future decisions. There is no action in isolation.

Why it matters

Agent harnesses often evaluate tool calls by their immediate output and ignore downstream state changes. This reframe makes second-order effects a first-class design concern for evaluators, simulators, and context tracking.

Generalization

Any decision in an interactive system should be modeled as an intervention on a closed loop, not as a one-way operation.

No decision you make exists in isolation.
Open source video
There are no side effects; there are simply effects you have not thought about yet.
Open source video
Mechanismmedium noveltystrong evidence

Stocks and flows provide a precise state-space abstraction: stocks accumulate over time and have memory, and they can only be changed by altering inflows or outflows. Confusing a stock-like quantity with a flow-like quantity (e.g., 'interest rate') leads to modeling errors.

Why it matters

In long-running agents, context and memory are stocks: they accumulate, retain history, and influence behavior. Designing them as managed stocks with explicit inflow/outflow policies is more effective than treating them as append-only logs.

Generalization

Any persistent quantity in a system—tokens, task queues, error budgets, reputation—should be represented as a stock manipulated through inflows and outflows, not as an instantaneous rate.

A stock is anything that accumulates over time and has memory.
Open source video
To change a stock, you must alter inflows or outflows.
Open source video
Interest rate is a price (stock-related), not a flow, despite its name.
Open source video
Failure Modehigh noveltymoderate evidence

Tight coupling and time delays between decisions and system-state changes produce overshooting and policy resistance when ignored. Any closed-loop agent that acts faster than the observable effect of its actions will tend to overcorrect and oscillate.

Why it matters

Agent planners that issue rapid corrective actions without waiting for observable feedback can cause overshooting, thrashing, and wasted tool calls. Delay-aware control—dampening, waiting for confirmation, or rate-limiting—is an architectural necessity.

Generalization

In any control loop, the delay between action and observable state must be smaller than or explicitly modeled against the loop's update rate, or the loop will become unstable.

Systems feature tight coupling and time delays between decisions and system state changes, leading to unintended consequences.
Open source video
Failing to account for delays leads to overshooting and policy resistance.
Open source video
Mechanismmedium noveltystrong evidence

Feedback loop polarity is mechanically analyzable: counting negative causal links in a loop determines whether it is reinforcing (amplifying) or balancing (goal-seeking). This gives engineers a practical method to predict whether a system interaction will amplify or stabilize.

Why it matters

Agent interaction graphs, retry loops, and multi-agent escalation paths can be classified as reinforcing or balancing. This classification predicts runaway growth, deadlock, or self-correction before simulation.

Generalization

Causal link polarity counting is a lightweight static analysis for feedback structures that can be automated or approximated on action graphs.

Counting negatives in a loop determines loop polarity (odd = balancing, even = reinforcing).
Open source video
Reinforcing loops amplify change; balancing loops seek goals or balance.
Open source video

Deep dives

4

Automated feedback loop polarity detection in agentic systems

Research question

How can LLM-generated causal graphs or causal discovery methods reliably infer reinforcing vs. balancing loop polarity from agent action/observation traces?

Why

Feedback loop polarity predicts whether an interaction will amplify or stabilize, enabling lightweight static analysis of agent interaction graphs before simulation.

Counting negatives in a loop determines loop polarity (odd = balancing, even = reinforcing).
Open source video
Reinforcing loops amplify change; balancing loops seek goals or balance.
Open source video
Source video

Stock-flow memory management for long-horizon agents

Research question

What are optimal inflow and outflow policies for context/memory as a stock in agents with bounded context windows?

Why

Context behaves as a stock with memory; managing it with explicit inflow and outflow policies prevents uncontrolled growth and preserves relevant information.

A stock is anything that accumulates over time and has memory.
Open source video
To change a stock, you must alter inflows or outflows.
Open source video
Source video

Delay-aware control for stable agent loops

Research question

How can agents measure action-effect delays and adapt their control rate to avoid overshooting and oscillation?

Why

Ignoring delays between decisions and observable state changes causes overcorrection, wasted tool calls, and policy resistance in autonomous agents.

Failing to account for delays leads to overshooting and policy resistance.
Open source video
Systems feature tight coupling and time delays between decisions and system state changes, leading to unintended consequences.
Open source video
Source video

Structural debugging: tracing agent failures to architecture rather than the model

Research question

What methods can systematically trace persistent agent failures back to structural causes such as information availability, decision rules, and incentives?

Why

The fundamental attribution error in AI debugging blames the model when persistent failures are symptoms of surrounding structure, so durable fixes require structural analysis.

Structure generates behavior.
Open source video
Dynamics emerge from the interaction of physics, information availability, and decision rules.
Open source video
Source video

Article ideas

4

There Are No Side Effects: Closing the Loop in Agent Architecture

Agent evaluations that ignore downstream state changes will misjudge tool use; every action is an intervention on a closed loop, and treating it as open-loop is a design flaw.

Angle

Using system dynamics' reframe of side effects to redesign agent harnesses and evaluators for second-order effects.

Source video

Stop Blaming the Model: Structural Debugging for Persistent Agent Failures

Persistent failures in agent systems are symptoms of structure—context visibility, decision rules, and incentives—not just model capability, and postmortems should trace those structures.

Angle

Applying the 'structure generates behavior' principle to AI agent postmortems, drawing on system dynamics.

Source video

The Context Window Is a Stock: Rethinking Memory as a Flow-Controlled Reservoir

Context and memory in long-running agents should be treated as a stock with explicit inflow and outflow policies, not an append-only log, to bound growth and preserve salient information.

Angle

Stock-and-flow modeling as a normative framework for context and memory management in LLM agents.

Source video

When Agents Overshoot: Delay-Driven Oscillation in Autonomous Loops

Agents that act faster than they can observe the effects of their actions will destabilize; delay-aware control is an architectural requirement for robust autonomy.

Angle

Modeling agent thrashing as a classic control-theoretic overshoot caused by hidden delays, with design patterns for damping.

Source video

Project ideas

4

LoopPolarity

beyond-evals

An LLM-based causal graph extractor can classify logged agent loops as reinforcing or balancing with >80% agreement with human coders.

Proof of concept

Collect traces from a multi-agent simulation, use an LLM to build causal graphs from the traces, compute loop polarity by counting negative links, and compare to human annotations.

Measurement

F1 score / agreement with human labels; predictive accuracy of runaway vs. stable outcomes.

Source video

StockFlow Memory Manager

new

Explicit outflow policies (summarization/eviction) in a long-horizon agent improve task success and bound context growth compared to append-only logs.

Proof of concept

Build an agent with a memory manager exposing inflow (add) and outflow (summarize/forget) knobs; benchmark on a long-horizon task with two policies: append-only vs. managed stock.

Measurement

Task completion rate, context utilization, and redundancy/irrelevant content ratio.

Source video

DelayAware Agent Controller

new

Inserting a configurable observation wait after actions, based on estimated environmental delay, reduces oscillatory behavior and wasted tool calls in interactive environments.

Proof of concept

In a simulated environment with adjustable delay, run an agent with and without delay dampening and compare corrective action counts and stability.

Measurement

Number of corrective actions, oscillation amplitude, and task success rate.

Source video

Mental Model Auditor

beyond-evals

Agents that state assumptions before acting and reconcile them with observed outcomes show lower policy resistance and fewer repeated mistakes.

Proof of concept

Instrument an agent to emit explicit predictions before each tool use; after outcomes, store mismatches and feed them back into subsequent decisions; compare to a baseline agent without this loop.

Measurement

Prediction calibration, repeated error rate, and decision revision frequency.

Source video

Architectural implications

4

Closed-loop feedback is fundamental: decisions alter system state, which feeds back into future decisions.

Before

Agent harness treats each tool call as an independent open-loop step and only captures the direct return value.

After

Agent harness explicitly closes the loop by feeding observable state changes and side effects back into context and using them for subsequent decisions.

Consequence

More accurate long-horizon behavior, but increased context load and the need to filter which state changes are relevant.

Source video

Stocks accumulate and have memory; they change only through inflows and outflows.

Before

Context window is managed as an append-only log that grows without a deliberate outflow policy.

After

Context and memory are managed as stocks with explicit inflow policies (what to add) and outflow policies (eviction, summarization, forgetting).

Consequence

Bounded memory and predictable context usage, but requires deliberate decisions about what is worth remembering.

Source video

Failing to account for delays leads to overshooting and policy resistance.

Before

Agent executes the next corrective action immediately after issuing a command, without waiting for the environment to respond.

After

Agent scheduler inserts observation waits or dampens repeated corrective actions until the effects of prior actions are visible.

Consequence

More stable control and fewer redundant tool calls, but slower loop iteration and the need for delay estimation.

Source video

Mental models filter information and shape actions, and misattributing behavior to individuals rather than structure is common.

Before

Debugging an agent failure stops at 'the model made a bad decision' without inspecting the structure that produced the decision.

After

The agent system records goals, assumptions, and decision rules, and postmortems trace behavior back to structure and incentives, not just model output.

Consequence

Better root cause analysis and more durable fixes, but requires more instrumentation and introspection from the agent.

Source video

Tradeoffs and failure modes

4

Closed-loop feedback vs. open-loop simplicity

Benefit

Captures 'effects you have not thought about yet' and enables self-correction.

Cost or risk

Complexity grows, and if delays are ignored the loop can overshoot and become unstable.

There are no side effects; there are simply effects you have not thought about yet.
Open source video
Source video

Mental models as filters

Benefit

Makes decisions tractable by simplifying the information space.

Cost or risk

Filters can hide systemic causes and lead to fundamental attribution error, blaming agents instead of structure.

Decision-makers operate based on mental models that filter information and shape actions.
Open source video
Source video

Model simplification

Benefit

Models make complex systems understandable and useful for policy design.

Cost or risk

All models are wrong; treating a model as accurate can produce confidently wrong predictions.

All models are wrong, but some models are useful.
Open source video
Source video

Reinforcing loops

Benefit

Amplify growth, learning, or adoption—useful for scaling what works.

Cost or risk

Without a balancing loop, reinforcing loops can generate runaway behavior or collapse.

Reinforcing loops amplify change; balancing loops seek goals or balance.
Open source video
Source video

Open questions

4

How can an agent system automatically infer the polarity of feedback loops from its own action and observation traces?

Why unresolved

Loop polarity is currently a manual causal diagramming exercise; traces contain correlation, not causal structure.

Research direction

Use causal discovery or LLM-generated causal graphs over logged actions and state changes to classify loops as reinforcing or balancing.

Source video

How should context and memory be designed as stocks with explicit inflow and outflow policies in long-running agents?

Why unresolved

Stocks and flows are an abstract modeling concept; there is no standard API for memory inflow (write) and outflow (forget/summarize) rates.

Research direction

Prototype memory managers that expose inflow/outflow controls and benchmark summarization/eviction policies against task performance.

Source video

How can an agent measure the delay between its own actions and observable environmental effects?

Why unresolved

Delays are usually not instrumented in agent harnesses, and environment causality is noisy.

Research direction

Add delay tracking metadata to tool calls and learn per-environment delay distributions; then test dampening or wait-for-effect policies.

Source video

Can an agent's internal 'mental model' be elicited and corrected to avoid policy resistance?

Why unresolved

Mental models are internal and not directly observable, making it hard to identify when they filter important feedback.

Research direction

Require agents to state assumptions and expected effects before acting, then compare against observed outcomes to detect model drift.

Source video

Key claims

8
comparativeVerification needed

System dynamics is control theory applied to social systems.

Evidence

System dynamics is control theory applied to social systems.

Question

What specific control-theoretic concepts transfer to social systems and which do not?

Source video
comparativeVerification needed

Social systems are much harder to control than physical systems.

Evidence

Social systems are much harder to control than physical systems.

Question

What are the properties of social systems that make control more difficult?

Source video
factualVerification needed

Counting negatives in a loop determines loop polarity: odd equals balancing, even equals reinforcing.

Evidence

Counting negatives in a loop determines loop polarity (odd = balancing, even = reinforcing).

Question

Does this rule hold universally for all signed causal loop diagrams?

Source video
factualVerification needed

A stock is anything that accumulates over time and has memory.

Evidence

A stock is anything that accumulates over time and has memory.

Question

How should one identify boundary cases where a quantity is partially a stock and partially a flow?

Source video
causalVerification needed

Failing to account for delays leads to overshooting and policy resistance.

Evidence

Failing to account for delays leads to overshooting and policy resistance.

Question

What empirical or simulated evidence supports this, and are there counterexamples?

Source video
opinionVerification not requested

There are no side effects; there are simply effects you have not thought about yet.

Evidence

There are no side effects; there are simply effects you have not thought about yet.

Source video
opinionVerification not requested

All models are wrong, but some models are useful.

Evidence

All models are wrong, but some models are useful.

Source video
causalVerification needed

When trying to fix a persistent problem, modifying the underlying structure and policies is more effective than blaming individuals.

Evidence

When trying to fix a persistent problem, modifying the underlying structure and policies is more effective than blaming individuals.

Question

What evidence in the video supports this beyond the oil price cycle example?

Source video

Connections

5