Agent versus workflow: choose the least complex reliable design

An agent chooses actions dynamically. A workflow follows known states and branches. If a refund form always needs validation, policy lookup, calculation and review, an open-ended planner adds variability without solving a real uncertainty.

Classify the task first

The agent controls lab evaluates two authored tasks.

from agent_cases import design_choice_case

result = design_choice_case()
assert result["workflow_count"] == 1
assert result["agent_count"] == 1
assert "least complex" in result["selection_rule"]
print(result["tasks"])

Form routing selects a workflow because its steps are known. Research across approved sources selects a bounded agent because tool choice depends on what is found. No LLM or framework executes.

Ask where dynamic choice adds value

Define inputs, allowed outputs, tool catalogue, branch conditions and failure handling. If you can draw the complete state machine, implement that as the baseline. Consider an agent for uncertain decomposition, source discovery or iterative diagnosis only when acceptance tests can bound it.

Compare task success, tool calls, latency, cost, recoverability and unauthorized-action rate. A more flexible design must outperform the workflow on the same cases. Preserve “workflow wins” as a valid result.

Keep controls in both designs

Dynamic planning does not remove schemas, permissions, timeouts, idempotency or approval. Bound steps and delegation, validate arguments and make termination explicit. Use deterministic code for calculations and authorization.

The Generative & Agentic AI course begins agentic work with a deterministic baseline and requires evidence for added coordination.

Exercise

Take one automation idea, draw its state machine and mark genuinely unknown decisions. Implement the workflow baseline, then add one bounded dynamic choice and compare both on 20 acceptance cases.

Continue learning

This article is part of the Agent workflows and state sequence. Use the neighbouring tasks when you need the prerequisite or the next application.

• Continue with Build a tool-calling loop with explicit termination.
• Then apply it in Validate tool arguments before execution.

Reference: LangGraph workflows and agents.