Scaffolding

In developmental psychology, scaffolding refers to the support structures provided by caregivers that enable children to accomplish tasks beyond their independent capability. As competence grows, scaffolding is gradually removed.

In cognitive science, the extended mind thesis proposes that cognition doesn’t stop at the skull—tools, notebooks, and environmental structures are part of our cognitive systems.

For AI agents, scaffolding encompasses both ideas: the external infrastructure that enables models to act as agents, extending their capabilities beyond raw language generation.

The Model is Not the Agent

A language model, by itself, cannot:

• Remember beyond its context window
• Take actions in the world
• Maintain state across sessions
• Recover from errors
• Manage multi-step processes

The scaffold provides all of this. It is the code, infrastructure, and tools that wrap a model to create an agent.

graph TB
  subgraph SCAFFOLD["SCAFFOLD<br/>action // persistence // reliability"]
      LM["LANGUAGE MODEL<br/>reasoning // language // knowledge"]

      M[Memory]
      T[Tools]
      S[State]

      P[Parsing]
      E[Error Handling]
      L[Loops]

      LM --- M
      LM --- T
      LM --- S

      M --- P
      T --- E
      S --- L
  end

  style SCAFFOLD fill:#0a0a0a,stroke:#00ff00,stroke-width:2px,color:#cccccc
  style LM fill:#0a0a0a,stroke:#00ff00,stroke-width:2px,color:#cccccc
  style M fill:#0a0a0a,stroke:#00ff00,stroke-width:1px,color:#cccccc
  style T fill:#0a0a0a,stroke:#00ff00,stroke-width:1px,color:#cccccc
  style S fill:#0a0a0a,stroke:#00ff00,stroke-width:1px,color:#cccccc
  style P fill:#0a0a0a,stroke:#333333,stroke-width:1px,color:#666666
  style E fill:#0a0a0a,stroke:#333333,stroke-width:1px,color:#666666
  style L fill:#0a0a0a,stroke:#333333,stroke-width:1px,color:#666666

Components of Scaffolding

1. The Agent Loop

The fundamental structure: code that repeatedly calls the model, parses outputs, executes actions, and feeds results back in.

while not done:
    response = model.generate(context)
    action = parse_action(response)
    result = execute(action)
    context = update_context(context, action, result)

This simple loop is the skeleton upon which all else hangs.

2. Tool Definitions

Schemas that define what actions are available, their parameters, and their effects. The model can only act through tools the scaffold provides.

3. Memory Systems

Mechanisms for persistence beyond the context window:

• Conversation history: Recent exchanges
• Summarization: Compressed representations of past interactions
• Vector stores: Semantic retrieval of relevant information
• External databases: Structured knowledge

4. State Management

Tracking where the agent is in a multi-step process:

• Current goals and subgoals
• Completed actions
• Pending tasks
• Error states

5. Error Handling

Recovery mechanisms when things go wrong:

• Retry logic for transient failures
• Fallback strategies
• Graceful degradation
• Human escalation

6. Output Parsing

Extracting structured actions from natural language:

• JSON parsing
• Regular expression matching
• Constrained generation
• Validation and correction

7. Guardrails

Safety constraints on agent behavior:

• Action filtering
• Rate limiting
• Permission systems
• Content moderation

The Extended Mind of Agents

The extended mind thesis becomes literal with AI agents. The model provides certain cognitive capabilities (language, reasoning, knowledge), while the scaffold provides others:

Neither is complete without the other. The agent is the combination.

Scaffolding Patterns

Different agent architectures emphasize different scaffolding patterns:

Minimal Scaffold

The model handles almost everything through prompting. Simple loop, few tools.

• Pros: Flexible, less code, model-adaptive
• Cons: Fragile, context-limited, hard to debug

Heavy Scaffold

Complex orchestration logic, many specialized components, model as a small part.

• Pros: Reliable, controllable, testable
• Cons: Rigid, expensive to build, model-agnostic

Hybrid Scaffold

Strategic allocation—scaffold handles reliability and persistence, model handles reasoning and adaptation.

• Pros: Best of both worlds
• Cons: Requires careful design

Scaffolding and Autonomy

The scaffold doesn’t just enable—it also constrains. Autonomy levels are implemented through scaffolding:

• Level 0-1: Scaffold mediates all actions, human in the loop
• Level 2: Scaffold filters actions, some require approval
• Level 3: Scaffold logs and monitors, intervention possible
• Level 4: Minimal scaffold, model drives

The scaffold is the control surface through which humans shape agent behavior.

Building Good Scaffolds

Effective scaffolding shares characteristics:

Transparency

The model can “see” the scaffold—understand its tools, memory, and constraints. Opacity leads to confusion.

Reliability

Failed API calls, malformed outputs, unexpected states—the scaffold handles these gracefully.

Observability

Humans can inspect scaffold state: what has the agent done? What is it trying to do? What went wrong?

Modularity

Components can be swapped, upgraded, or removed without rewriting everything.

Appropriate Coupling

The scaffold should be tight enough to ensure reliability but loose enough to accommodate model updates.

The Future of Scaffolding

Current scaffolding is largely hand-crafted. Future directions include:

• Learned scaffolds: Models that generate their own tool definitions and orchestration
• Self-modifying scaffolds: Agents that improve their own infrastructure
• Scaffold marketplaces: Reusable components and patterns
• Scaffold-free agents: Models capable enough to internalize scaffolding functions

The Scaffolding Paradox

Here’s the tension: scaffolding enables agents but also limits them. The more structured the scaffold, the more reliable but less flexible the agent.

This mirrors human development. Children need scaffolding to learn, but eventually, the scaffolding must be removed for full maturity. Whether AI agents will ever mature beyond their scaffolds—or whether they’ll always be hybrid systems—remains an open question.

See Also

• Tool Use — the action component of scaffolding
• The Agent Loop — the fundamental scaffold structure
• Autonomy Levels — how scaffolding implements oversight
• Memory Systems — the persistence component of scaffolding