4 A Capability Model

1. Why the 4 A's Exist

The 4 A's provide a classification layer for deciding how much reasoning, execution freedom, and human involvement a Site capability requires.

They are not four unrelated products and they are not a strict four-step maturity ladder.

flowchart LR
    A[["Assistant<br/>User-led help"]]:::model -->|"more dynamic decision freedom"| G[["Agent<br/>Goal-led action"]]:::model
    G -->|"bounded independent continuation"| AU[["Autonomous<br/>Policy-bounded operation"]]:::model
    W[["Automation<br/>Predefined deterministic execution"]]:::process -.->|"can support"| A
    W -.->|"can be invoked by"| G
    W -.->|"can be orchestrated by"| AU

    classDef process fill:#EEF2F7,stroke:#475569,stroke-width:1px,color:#0B1220
    classDef model fill:#EDE9FE,stroke:#6D28D9,stroke-width:1px,color:#0B1220

2. Canonical Definitions

Assistant

An Assistant primarily supports a user. It may retrieve knowledge, generate content, summarize, classify, or invoke narrowly controlled tools, but the user remains the primary driver of the interaction.

Agent

An Agent receives a goal and can dynamically select steps or tools, observe results, and continue execution until it reaches a completion condition or a control boundary.

Automation

Automation executes a predefined process. It can include conditions, schedules, events, APIs, scripts, jobs, and even model calls, but the process structure remains defined by the designer rather than dynamically planned by an agent.

Autonomous

Autonomous describes an operating mode where the platform can continue toward an objective across time with limited human intervention, including re-planning and deciding what to do next within policy, permission, cost, and safety limits.

3. Comparison

Dimension	Assistant	Agent	Automation	Autonomous
Primary input	User request	Goal/task	Trigger/event/schedule	Objective/condition
Step selection	Mostly fixed	Dynamic	Predefined	Dynamic and continuing
Tool selection	Limited/configured	Dynamic within allowlist	Predefined	Dynamic within policy
Human involvement	High	Medium	Low	Exception-based
Planning	Minimal	Yes	No	Yes, repeated
Long-running	Optional	Optional	Common	Common
Deterministic	Low-medium	Low-medium	High	Low-medium
Approval	Usually user driven	Often	Configurable	Policy-based
Best runtime	Chat/direct API	Agent SDK/orchestrator	n8n/workflow/code	Agent + automation + policy loop

4. Classification Questions

Use the following order when classifying a requirement:

1. Is a human actively asking for help? If yes, start with Assistant.
2. Must the system decide which steps/tools are necessary? If yes, Agent is involved.
3. Is the process sequence already known? If yes, use Automation for that sequence.
4. Must the capability continue operating after the initiating interaction ends? If yes, evaluate Autonomous mode.
5. Can actions affect external systems? If yes, define approval, permission, rollback, and audit boundaries regardless of type.

5. Composition Pattern

A realistic Site capability may combine all four.

flowchart TD
    AO[["Autonomous Asset Health Operation"]]:::model --> AG[["Agent"]]:::model
    AG --> A1["Analyze telemetry and context"]:::process
    AG --> A2["Determine required action"]:::process
    AO --> AT[["Automation"]]:::process
    AT --> T1["Create work order"]:::process
    AT --> T2["Send notification"]:::process
    AO --> AS[["Assistant"]]:::model
    AS --> S1(["Explain decision to engineer"]):::output
    AS --> S2(["Answer follow-up questions"]):::output

    classDef process fill:#EEF2F7,stroke:#475569,stroke-width:1px,color:#0B1220
    classDef model fill:#EDE9FE,stroke:#6D28D9,stroke-width:1px,color:#0B1220
    classDef output fill:#DCFCE7,stroke:#15803D,stroke-width:1px,color:#0B1220

6. Classification Is Separate from Implementation

The same logical type can be implemented using different products.

Logical Type	Possible Implementations
Assistant	Direct model SDK, Open WebUI, Flowise Assistant, custom Site UI
Agent	OpenAI Agents SDK, Microsoft 365 Agents SDK, Flowise Agentflow, custom runtime
Automation	n8n, code workflow, event action, scheduled job
Autonomous	Agent runtime + policy + scheduler/event loop + automation/tool execution

7. Anti-Patterns

Avoid the following classifications:

• Calling a chatbot an agent only because it uses an LLM.
• Calling a scheduled workflow autonomous because it runs without a user.
• Calling an n8n workflow an agent when all steps are predefined.
• Calling an agent autonomous when it cannot continue beyond a single delegated run.
• Embedding a third-party AI UI and treating the embedded UI itself as the capability definition.

The Site registry should describe the capability, while the implementation reference describes how it is executed.