Since 1994

Workflow or agent?

A decision framework before you architect anything · Rex Black, Inc.

Most "agents"

are workflows

that overshot.


Autonomy is expensive in three measurable ways.

Debug cost compounds with run length.
Observability cost is a prerequisite, not an upgrade.
Drift cost of a system that picks its own next step is recurring.

REX BLACK, INC. · WORKFLOW OR AGENT
Definitions

Two shapes.

Both valuable.

Workflow

Deterministic pipeline.
LLM steps are local: extract, classify, summarize.
Control flow lives in code.

Agent

LLM picks the next step.
Has tools. Invokes them in an order it chooses.
Control flow lives in the model.

REX BLACK, INC. · WORKFLOW OR AGENT
01 / 04

Can the steps be

enumerated in advance?

If the solution path is knowable at design time, the workflow wins.

Agents earn their keep when the path is genuinely variable. Most enterprise tasks are not.

Test. Write the solution path on a whiteboard. If you can, it is a workflow.

REX BLACK, INC. · WORKFLOW OR AGENT
02 / 04

What is the

blast radius of

a wrong next step?

Agents fail distinctively: they do the wrong thing efficiently.

A workflow dead-ends with an exception.
An agent dead-ends, tries another tool, writes to something it should not have touched.

Low blast radius. Autonomy is cheap.
High blast radius. Autonomy is a contingent liability priced in incidents.

REX BLACK, INC. · WORKFLOW OR AGENT
03 / 04

How observable

is the decision trace?

Workflow

Observable by construction. Steps are in code. Logs are enough.

Agent

Requires a separate observability layer: trace capture, tool-call replay, decision-point annotation.

Build it before the agent.

REX BLACK, INC. · WORKFLOW OR AGENT
04 / 04

Unit economics

at full volume.

At 100/day, the cost difference is noise.

At 100,000 requests/day, agents cost 2 to 4x more per decision and the surface is harder to bound.


Workflow first is not ideology. It is an economic argument, with the chart to prove it.

REX BLACK, INC. · WORKFLOW OR AGENT

The earned-autonomy

principle.


Build the workflow. Accept 85 to 95 percent coverage.
Instrument the failures. Capture inputs, intermediates, low-confidence paths.
Look at the distribution. Clustered? Solve in the workflow. Long-tail? Evidence for an agent layer.
Add a bounded agent, not a replacement. Narrow allowlist. Explicit territory.

REX BLACK, INC. · WORKFLOW OR AGENT
Worked examples

Three real decisions.

Doc processing · 50K/month

Workflow. Fields enumerable. Blast radius bounded. If a doc type fails, narrow agent layer for that type.

Support triage · 8 queues

Workflow. LLM as classifier. Sequence is fixed: classify, score, route, log. No agent needed.

IT ops remediation, 15 system types

Agent, cautiously. Tool allowlist, step caps, mandatory human approval on writes, long-running eval on traces.

REX BLACK, INC. · WORKFLOW OR AGENT

Failure modes

by path.


Workflow

LLM step swallows the error.
Pipeline ossifies past 92%.
Prompt file becomes unversioned product.

Agent

Observability deferred → rewrite in 6 months.
Unbounded retries → linear cost in failure rate.
Policy drift → system prompt edits that nobody audited.

REX BLACK, INC. · WORKFLOW OR AGENT

This week.


Re-scope one in-flight "agent" project as a workflow. Walk through the four questions.
Audit observability on any production agent. If "why did it do X" is unanswerable, the layer is insufficient.
Name one workflow candidate for earned autonomy. Look at the failure distribution.

REX BLACK, INC. · WORKFLOW OR AGENT
Rex Black, Inc.

Workflow first.

Agent earned.

rexblack.com/resources/writing/workflow-or-agent