Prompt vs Workflow vs Agent

Educational prototype — not intended to make real credit decisions or evaluate real consumers.
An Inexacto Lab experiment

Compare accuracy, latency, cost, control, and autonomy on the same synthetic credit-review task.

Experiment setup

Prompt

A single LLM call with no tools and no iteration. The model reads the policy and application once and produces a recommendation directly.

Complexity
Low
Autonomy
Low

Workflow

A deterministic, code-controlled pipeline: extract application → extract policy → validate → analyze risk → recommend. The LLM executes each stage; code controls the sequence.

Complexity
Medium
Autonomy
Medium

Agent

Claude with tool use and autonomy over its own execution path (read_application, search_policy, check_rule, calculate_metric, request_human_review), bounded by a step limit and a timeout.

Complexity
High
Autonomy
High

Credit application

The case all three architectures will analyze below.

The application JSON is invalid. Switch to raw JSON to fix it.

Credit policy

The rules used to evaluate this case.