Agentic Workflow

An agentic workflow is a process where an AI model chooses some of the next steps itself, based on what it finds, instead of following a fixed path every time. A fixed workflow says: when a form arrives, copy these fields to the CRM and send this email. An agentic workflow might read a messy email, decide which of five tools to call, check the result and decide what to do next.

Workflows and agents are a spectrum

Anthropic's engineering guide on building effective agents separates workflows, where code sets the path and models fill in steps, from agents, where the model directs its own process and tool use. Most business systems we build sit in between: a fixed outer path with one or two steps where a model reads, classifies or drafts.

Use the simplest option that works

Anthropic's guide recommends starting with simple patterns and adding agent behavior only when it clearly improves results, because agents trade cost and speed for flexibility. Fixed steps are cheaper to run, easier to test and give the same output for the same input. Model-driven steps handle variety but need evaluation, limits and monitoring.

What makes an agentic workflow safe enough to run

is the boundary around it: a short list of allowed tools, read-only access where possible, human approval for actions with real consequences, a cap on steps and spend, and a log of every decision. Without those, a workflow that can choose its own steps can also choose wrong ones.

A business example

a freight team's inbox gets carrier emails in many formats. An agentic step reads each one, pulls the container number and new date, and updates the shipment. A fixed step then checks the date against free time and alerts a person if a charge is close. The model handles variety; the rules handle money.