Agentic workflows

Overview

• An agentic workflow allows a model to choose some steps based on observations.
• The surrounding application limits the actions, resources, and impact of those choices.
• It is useful for variable tasks but offers less predictable paths and call counts.

Why this matters

A user may omit information or ask a question requiring several possible information sources.

An everyday example

A support worker chooses whether to ask a question, read a handbook, or contact another team.

A technical example

user → agent → choose search/tool/clarify/escalate → observe → continue.

How it works

1. Define goal and allowed choices
2. Let model select
3. Enforce permissions
4. Observe
5. Check progress/budget
6. Finish or escalate.

Architecture and flow

User → agent → decide → search / tool / ask / escalate / bounded act.

The insurance assistant example

Decide which document checklist to retrieve after learning the claim type.

When to use it

Adaptive information gathering and variable low-risk tasks.

When to avoid it

Do not let a model invent deductible rules, bypass approvals, or retry writes based on conversational optimism.

Interview answer

“Agentic workflows place bounded model choices inside an application-controlled process; they are useful where next steps are variable.”

Benefits, trade-offs, and production considerations

Measure completed tasks, unnecessary calls, loop rate, escalation quality, and cost distribution. Trace chosen paths, not only final text.