Agents and workflows

ConceptsChapter 3 of 142 min read

The difference between a prompt, a workflow, and an agent comes down to one question: who decides what to do next?

Prompts, workflows, and agents#

Single prompt: Input text goes in, output text comes back. One call, zero branching.

Workflow (e.g., RAG): Code executes predetermined steps in a fixed sequence.

Agent: An LLM evaluates a goal, inspects tool outputs, and dynamically chooses the next action in a loop.

Drawing diagram…

A workflow executes a predetermined sequence; an agent loops dynamically.

Architecture	Control flow	LLM calls per run	Adaptability	Cost & latency
Prompt	User	1	None	Lowest
Workflow	Application code	Fixed	Handles anticipated edge cases	Predictable
Agent	LLM	Variable	Handles unexpected formats & outputs	Higher

Use an agent when the steps required to complete a task cannot be predicted upfront:

Handling unstructured inputs: Contract formats vary widely across vendors—Word files, scans, custom clauses.

Dynamic recovery: If an initial search returns incomplete terms, an agent can re-query with modified parameters instead of failing.

Task specialization: Complex jobs are divided among specialized personas (e.g., term extraction, risk analysis, fact-checking).

Agents introduce non-deterministic execution paths, higher token costs, and longer latency.

Do next steps depend on unpredictable runtime findings? Build an autonomous agent.