“Agentic” describes a family of implementations, rather than a standardized autonomy level. Anthropic distinguishes workflows with predefined code paths from agents that dynamically direct their own process. Other teams use “agentic workflow” for both or for a mixture. AI workflow is the broader term, which also covers fixed paths where the model never chooses the next step.
Example: a ticket-handling workflow can classify the request, retrieve a policy, let a model choose a follow-up search, then require an application check before any account change. Some transitions are fixed; one is model-directed. That is more useful to know than whether the product is called an agent.
For design reviews, draw the actual control path and name who chooses each transition. Record stopping conditions, failure handling, and actions that require additional authorization. Structure can make behavior easier to inspect, but does not by itself establish reliability.
What makes an AI agent? covers the action loop, and how to deploy agents safely connects that loop to execution controls and evaluation.
Sources
- Anthropic: Building effective agents — Defines workflows and agents by who controls the process, rather than by a fixed autonomy ladder.
Go deeper
From the conversation
-
Got Agents? Agentic Workflows & Architecture | Weaviate, Unstructured & CrewAI -
11 Cameras, Dozens of Mics: The AI That Reads the Room | Tormod Ree, Neat -
The Making of Gemini 2.0: DeepMind's Approach to AI Development and Deployment | Logan Kilpatrick -
Practical Lessons for GenAI Evals | Chip Huyen & Vivienne Zhang -
Hallucinations Are a Data Architecture Problem | Sudhir Hasbe, Neo4j -
Mastering Multi-Agent Systems | MongoDB’s Mikiko Chandrasekhar -
Parenting Your AI Agents for Best Results | Cisco's Jeetu Patel -
The First Fully Autonomous AI Attack is Coming | Kris Lovejoy, Kyndryl