Workflow Architecture Classification

1. Purpose and scope

This framework is used to describe the control architecture a workflow requires. It separates execution shape, reasoning needs, system access, external consequences and human authority so that a workflow is not labeled “agentic” merely because it touches software or performs a write.

2. Definitions

Pathway

A rule-based workflow with predefined steps, routing and conditions.

Agent

An AI actor that interprets context and can choose how to progress within its configured capabilities and controls.

Governed Mission

A bounded objective in which an Agent operates under explicit scope, policy, permissions and human-control requirements.

Thinkway

A reusable reasoning and state method inside an Agent. It structures how the Agent works; it is not a classifier output and does not itself grant authority.

3. Six classification dimensions

1. Predictability. How much of the execution can be expressed as predefined steps, conditions and expected branches?
2. Interpretation. How much contextual interpretation or judgment is required to determine the next useful step?
3. System access. Which systems must be read or written, and which configured interfaces make that access possible?
4. External effect. Can execution create a consequential effect outside the reasoning process, such as changing a record, sending a message, scheduling an appointment or triggering another system?
5. Human authority. Which decisions, approvals and exceptions remain with a human, and which actions are explicitly authorized for the system?
6. Execution shape. Is the workflow predominantly deterministic, reasoning-driven, a governed agent mission, or a deliberate combination?

System access, write access, external effect and human approval change the control architecture. None of them alone proves that an AI Agent is required.

4. Deterministic decision process

1. Record evidence for all six dimensions before selecting an output.
2. Determine whether the required execution can be represented by a predefined workflow, requires contextual Agent reasoning, or combines both shapes.
3. Model system reads, writes, external effects and human approvals as control-architecture requirements rather than using any one of them as proof of agentic execution.
4. Apply the established output definitions below. When the available evidence does not support one classification cleanly, use NEEDS_CLARIFICATION instead of guessing.

Classification outputs

10. Thinkway as a supporting reasoning/state method

A Thinkway lives inside an Agent. It provides a reusable structure for reasoning or work state so professional methods can be repeated consistently. It is not an alternative to PATHWAY, AGENT, AGENT_IN_GOVERNED_MISSION, HYBRID or NEEDS_CLARIFICATION.

Pathways automate the process. Thinkways structure how the agent works.

11. Authority is independent of autonomy

How independently an Agent reasons does not determine what it is authorized to do. Tool access, permissions, approval requirements and policy remain separate control decisions.

12. System access and external consequence

Read access, write access and consequential effects must be represented explicitly in the architecture. A deterministic Pathway can perform configured writes; an Agent can also be limited to reads. The presence of a write or external effect does not by itself select the Agent classification.

13. Human-control boundary

Identify which actions require human approval, which exceptions remain staff-only and which effects are authorized without intervention. Human approval changes the control path; it does not by itself determine the execution shape.

14. Worked examples

Example · PATHWAY

A request is validated against fixed fields, routed by explicit rules and written to a configured system only after a predefined approval condition. The execution shape remains deterministic.

Example · AGENT

A support Agent interprets an open-ended customer question, retrieves permitted context and decides which configured response or tool step is appropriate. Its authority is still bounded separately from its reasoning.

Example · AGENT_IN_GOVERNED_MISSION

An Agent is assigned a bounded research-and-resolution objective with explicit permitted tools, policy checks, approval requirements and a defined completion condition.

Example · HYBRID

A deterministic intake and authorization flow hands one ambiguous interpretation step to an Agent, then returns the result to predefined execution and approval steps.

Example · NEEDS_CLARIFICATION

A proposed workflow says it should “handle bookings automatically” but does not define the target system, permitted writes, exception ownership or whether contextual interpretation is required. Classification remains open until those requirements are known.

15. Non-goals / limitations

This framework does not prove that a workflow is safe, compliant, economical, production-ready or technically integrated. Those questions require their own evidence and controls.

This is Createting's architecture classification framework. It is not an industry-standard AI maturity score and not a deployment-readiness certification.

16. Classifier version / change history

1.0 · 2026-09-15. Initial technical-reference publication candidate for the accepted Workflow Architecture Classification framework. Draft state; no production publication in this closure.

17. Interactive Assessment

Use the interactive Assessment when you want to apply the same architecture questions to a specific workflow.