Workflow vs. agent: when BPMN-style orchestration beats dynamic reasoning

Process-automation teams and AI teams are currently building the same systems with opposite defaults: everything-is-a-workflow versus everything-is-an-agent. Both defaults fail expensively. Here’s the decision framework.

What each one actually is

A workflow engine (BPMN-style orchestration) executes a process defined at design time: steps, branches, retries, timeouts, human tasks — all enumerated before the first execution. Determinism is the product.

An agent decides its next step at runtime by reasoning over the current state. Adaptivity is the product — and non-determinism is the cost, paid on every single execution.

The bright-line principle

If you can draw the happy path and its exceptions on a whiteboard, it’s a workflow. If the path depends on understanding content you haven’t seen yet, that step is an agent.

Note the phrasing: that step. The unit of decision is the step, not the system.

The comparison

Dimension	Workflow	Agent
Behavior across runs	Identical, replayable	Varies with input phrasing
Failure mode	Stuck token, visible in the engine	Plausible wrong action
Auditability	Free — the diagram is the doc	Must be engineered (trajectory logs)
Change process	Redeploy a definition	Prompt/model change → re-evaluate
Cost per execution	Compute, ~fixed	Tokens, variable and content-dependent
Handles novelty	Only what was modeled	Yes — that’s the point

The pattern that wins: agent steps inside workflow rails

The strongest enterprise architecture in practice is a deterministic workflow spine with agent-shaped holes:

The workflow owns sequencing, SLAs, retries, human-approval gates, and the system-of-record writes.
Agent steps sit inside individual workflow tasks: “interpret this document,” “draft the response,” “triage this case” — bounded work with a defined input and a schema-validated output.
The agent’s output re-enters the workflow through validation: schema checks, confidence thresholds, and a human-review branch below the threshold.

The workflow never asks the agent “what should happen next in the process?” — that question belongs to the diagram. The agent never gets to write to a system of record directly — that mutation belongs to a workflow step that can be audited and compensated.

When to break the rule

Fully agentic control flow earns its keep when the space of situations is genuinely unenumerable — open-ended research, exploratory coding, incident diagnosis. Signs you’ve misclassified: your “agent” follows the same trajectory 95% of the time (that’s a workflow wearing a costume), or your “workflow” has grown a branch for every customer email variant ever received (that’s an agent suppressed by management).