What is an agentic workflow?

An agentic workflow is a multi-step process in which an LLM does part of the deciding: it plans steps, picks tools, routes work or checks its own output. The term has no single definition. Model providers and framework docs split it in two: workflows, where your code fixes the path, and agents, where the model chooses the path as it runs.

For the broader idea, see what is agentic AI and what are AI agents.

Two meanings of the term

Anthropic’s “Building effective agents” (December 2024) drew the line that LangGraph’s docs also use. Anthropic notes that much of the tooling the post describes has changed since, but the definitions are unchanged:

• Workflows are “systems where LLMs and tools are orchestrated through predefined code paths.”
• Agents are “systems where LLMs dynamically direct their own processes and tool usage, maintaining control over how they accomplish tasks.”

LangGraph’s docs use the same split: workflows “have predetermined code paths,” while agents “are dynamic and define their own processes and tool usage” (LangChain docs, read 2026-10-30).

Many vendor guides use the word differently. Databricks defines it as processes that “plan, execute, and refine multi-step tasks,” where “the agent determines its next step based on context and intermediate results” (Databricks). Neo4j says an agentic workflow “decides the script as it runs” (Neo4j). Both put run-time decisions by the model at the center, which is how Anthropic defines an agent.

So when a vendor says “agentic workflow”, ask who decides the next step: your code or the model. That one question tells you how predictable the system is and how you will test it.

Who decides the next step

	Workflow (code decides)	Agent (model decides)
Path	Fixed in code	Chosen at run time
Number of steps	Known up front	Not known up front
Anthropic’s advice	Offers “predictability and consistency for well-defined tasks”	Better “when flexibility and model-driven decision-making are needed at scale”
Framework examples	Sequential, Loop and Parallel workflow agents in Google ADK; chaining and parallel execution in the OpenAI Agents SDK	Handoffs and agents as tools in the OpenAI Agents SDK; tool-calling agents in LangGraph

Sources for the table: Anthropic, Google ADK, OpenAI Agents SDK, LangGraph, all read 2026-10-30.

The line is not always sharp. Anthropic’s orchestrator-workers pattern has “a central LLM” that “dynamically breaks down tasks,” yet Anthropic still lists it as a workflow. Treat the two as ends of a range.

The five workflow patterns

Anthropic names five patterns for workflows. LangGraph documents the same five.

Pattern	What it does (Anthropic’s wording)
Prompt chaining	“decomposes a task into a sequence of steps, where each LLM call processes the output of the previous one”
Routing	“classifies an input and directs it to a specialized followup task”
Parallelization	LLMs “work simultaneously on a task and have their outputs aggregated programmatically”
Orchestrator-workers	“a central LLM dynamically breaks down tasks, delegates them to worker LLMs, and synthesizes their results”
Evaluator-optimizer	“one LLM call generates a response while another provides evaluation and feedback in a loop”

Source: Anthropic, read 2026-10-30. Neo4j describes agentic workflows with a different list: planning, tool use, reflection and orchestration.

Orchestrating with code or with the model

The OpenAI Agents SDK docs describe the same choice for multi-agent systems. With LLM orchestration, you are “allowing the LLM to make decisions”, using agents as tools or handoffs. With code orchestration, you determine “the flow of agents via your code” using structured outputs, chaining, an evaluator loop or parallel execution. The docs say code orchestration makes speed, cost and performance more predictable (OpenAI, read 2026-10-30).

Google’s Agent Development Kit has a code-side version: workflow agents (Sequential, Loop, Parallel) control the order of sub-agents “without consulting an AI model for assistance with the orchestration,” so execution is deterministic. The docs add that in ADK 2.0 for Python and Go, graph-based and dynamic workflows supersede these template workflows (Google ADK, read 2026-10-30).

Choosing between them

Anthropic’s guidance is to find “the simplest solution possible, and only increasing complexity when needed.” Use a workflow when the steps are known and a fixed path will do. Reach for an agent when it is hard or impossible to predict the number of steps a task needs.

Testing and tracing a workflow

A fixed workflow can be tested step by step, because each LLM call has a known place in the path. An agent needs the whole run judged, since the sequence differs between runs. In both cases, record every model call and tool call as one trace. The agent observability page covers what a trace should contain.

HoneyHive traces workflows and agents built with frameworks such as LangGraph; see the HoneyHive LangGraph integration docs for setup.

FAQ

What is the difference between a workflow and an agent?

In Anthropic's definitions, workflows orchestrate LLMs and tools "through predefined code paths", while agents are systems where LLMs "dynamically direct their own processes and tool usage". LangGraph's docs draw the same line.

Is an agentic workflow the same as an AI agent?

Often, in practice. Several vendor guides, such as Databricks and Neo4j, use "agentic workflow" for a process where the next step is decided at run time, which Anthropic would call an agent. Others use it for any LLM-based multi-step flow. Check which meaning a source uses. See what are AI agents.

When should I use a workflow instead of an agent?

Anthropic recommends finding "the simplest solution possible" and adding complexity only when needed. Workflows suit well-defined tasks. Agents suit problems where you cannot predict the number of steps.

How do I debug an agentic workflow?

Record each model call, tool call and handoff as one trace for the whole run. The agent observability page covers what to capture.