The core question
Every week, a founder asks us the same thing: should this be an agent or a workflow? The answer isn't about hype — it's about how much variability exists in the task.
Workflows excel when steps are known, inputs are structured, and exceptions are rare. Agents excel when context changes, decisions require judgment, and the right action depends on unstructured information.
Choose a workflow when
The process has fixed steps that rarely change. Data arrives in predictable formats. Errors should halt execution, not trigger improvisation. Compliance requires deterministic audit trails.
Examples: syncing CRM records on a schedule, generating a weekly report from a database, routing form submissions to the right Slack channel based on a dropdown field.
Choose an agent when
The task requires reading unstructured input — emails, PDFs, chat messages, call transcripts. Multiple tools may need to be called in an order that depends on context. Human-like judgment is needed, but within guardrails.
Examples: qualifying inbound leads from free-text inquiries, reviewing contracts against a policy library, triaging support tickets that don't fit clean categories.
The hybrid approach we recommend
Most production systems we ship are hybrids. A workflow handles the reliable backbone — triggers, data sync, notifications. An agent handles the variable decision point in the middle.
This gives you the predictability operators need with the flexibility models provide. It also makes debugging easier: when something breaks, you know whether to check the pipeline or the reasoning layer.
What to do next
Map your target process on paper. Circle every step that requires reading unstructured input or making a contextual decision. Those circles are your agent candidates. Everything else is workflow territory.
If you're unsure, start with a workflow for the predictable 80% and add an agent only where manual review is eating your team's time today.