WHITEBOXX / GLOSSARY

Agentic workflows: AI that
carries out the steps.

An agentic workflow is a business process in which one or more AI agents carry out some of the steps themselves. They work out how to complete each step, use tools such as searches and calculations, and check their results, within limits that people set.

By Gautam Mehra · Published · Updated

What is an agentic workflow?

In a traditional automated workflow, every step and every branch is written in advance. In an agentic workflow, some steps are given to an AI agent as a goal, such as “find the latest filings for these suppliers and flag any change in ownership”. The agent works out the searches and checks needed to reach it.

Most enterprise agentic workflows mix the two. Fixed steps handle the parts that must always happen the same way, such as approvals and exports. Agents handle the parts that need judgement, such as reading documents or matching records that do not agree.

COMPARISON

Agentic workflow vs workflow automation.

COMPARISON
AspectWorkflow automationAgentic workflow
StepsWritten in advancePartly decided by the agent while it runs
InputsStructured and predictableCan include documents, messages and messy records
When something unexpected happensThe run stops or follows a fixed ruleThe agent can try another approach, within its limits
How you check itTest each ruleTest results against many examples and keep a record of each run
Good forHigh-volume, repeatable tasksTasks that need reading, comparison or judgement

Why do agentic workflows fail?

Andrew Ng popularised the term in 2024 when he described four design patterns for agents: reflection, tool use, planning and collaboration between several agents. The patterns work well in demos. In production, most failures come from the parts around the agent: data it cannot reach, tools without clear limits and results nobody checks.

Gartner expects more than 40% of agentic AI projects to be cancelled by the end of 2027, mainly because of rising costs, unclear value or weak risk controls.

DESIGN

How to design an agentic workflow.

  1. 01

    Start with one complete task

    Pick a task with a clear input, a clear result and a person who uses it.

  2. 02

    Mark the fixed steps

    Approvals, exports and notifications should not be left to an agent.

  3. 03

    Give each agent step a goal and limits

    Say which data and tools it may use and how much it may spend.

  4. 04

    Decide who reviews what

    Agree which results a person must approve before they are used.

  5. 05

    Record every run

    Keep the inputs, steps and outputs so a result can be checked later.

How Evidence OS runs agentic workflows.

Evidence OS workflows combine fixed steps with execution-agent assignments. Each assignment has its own tools and budget, and each run is recorded with its inputs, step progress and deliverables.

See how workflows run