Agents Versus Workflows

Fact-check status

Reviewed claim by claim on 2026-09-02 against the primary or official sources identified inline as S1 and S2. Recommendations are design guidance unless a cited standard makes them normative.

Key jargon

Term	Plain-language meaning
Workflow	A mostly predefined path through known steps. S1
Agent	A system in which a model selects some next actions based on observed state. S1
Autonomy	The degree of freedom to choose goals, steps, tools, or timing. S1
Determinism	The extent to which the same input follows the same path and produces the same result. S1

Key concepts

Use a workflow when the path is known; introduce agentic choice only where variability creates measurable value. S1
Autonomy is a design variable, not a binary product category. S1

Concept map

flowchart LR
    A["Known path favors workflow"] --> B["Unknown choice invokes agent"]
    B --> C["Harness bounds autonomy"]
    C --> D["Evaluate outcome and path"]

Diagram semantics verified against S1 and S2 on 2026-09-02.

A workflow follows code-defined paths. An agent uses a model to choose some next steps or tools based on current state and observations. S1

Use a workflow when	Consider an agent when S1
Steps and decisions are known	Paths cannot be enumerated economically S1
Determinism and auditability dominate	Flexible interpretation creates real value S1
Failure cost is high	Actions remain bounded and observable S1
A rules engine expresses policy	Tool selection needs semantic judgment S1

Many effective systems combine both: deterministic outer workflow, model-assisted decisions inside narrow nodes, deterministic validation, and human approval at consequential transitions. S1

Exercise

Classify invoice extraction, travel research, database migration, incident triage, and deleting unused cloud resources. Identify which decisions should remain deterministic.

Principle

Start with the simplest architecture that works. Autonomy adds state space, cost, latency, and failure modes. S1