On this page
Learning objectives
- Define workflow and agent in operational terms
- Match autonomy to task uncertainty and action risk
- Choose the simplest architecture that can meet the outcome
Prerequisites
Basic understanding of API-based software
Answer first
Canonical definition
A workflow follows a predefined control path, while an AI agent uses a model to choose the next action from an allowed set based on observations, state, and a termination condition.
Operating context
Why it matters
Most business automation does not need open-ended planning. Deterministic orchestration is easier to test, operate, and audit.
Agents earn their complexity when the path varies, inputs are unstructured, and a useful decision cannot be enumerated in advance.
Method
Step by step
- 01
Map the path
List known steps, branches, tools, exceptions, and the information needed at each decision.
- 02
Locate uncertainty
Identify which decisions require interpretation rather than a stable rule or lookup.
- 03
Score action risk
Separate read-only analysis, reversible writes, external communication, and high-impact irreversible actions.
- 04
Choose the minimum autonomy
Use a workflow by default, add a model-assisted decision where needed, and introduce an agent loop only when repeated adaptive action is necessary.
Hands-on lab
Run an architecture decision review
Compare a rules workflow, model-assisted workflow, and agent for the same business task.
Deliverable
A decision record covering uncertainty, tools, state, failure impact, evaluation method, and operating cost.
What breaks
Common failure modes
- F1Using an agent for a stable sequence of API calls
- F2Hiding business rules inside a long system prompt
- F3Granting broad tool access before proving the task boundary
Beyond the demo
Production notes
- Every agent needs a maximum step count, time or cost budget, and explicit termination states.
- Keep deterministic validation and policy checks outside the model whenever possible.
Further reading
Sources
Related Tenten resources