Agentic workflow
Overview
An agentic workflow is a task-execution pattern in which a language model acts not only as a generator of text but as an active decision-maker: it determines the steps needed to accomplish a goal, calls tools to gather information or take actions, evaluates intermediate results, and adjusts its plan based on feedback — repeating this loop until the task is complete or a termination condition is reached.
The term contrasts with single-turn generation, where the model produces one response and the task ends. In an agentic workflow the model is embedded in a loop that may span many steps, tool invocations, and model calls.
The vocabulary in this area is unsettled: agentic workflow, agent loop, AI agent, and autonomous agent are used with overlapping meanings. This page uses agentic workflow specifically for the execution pattern, reserving AI agent for the system that runs the workflow.
Structural elements
A typical agentic workflow includes:
| Element | Description |
|---|---|
| Goal specification | The task or objective given to the agent (user instruction or programmatic input) |
| Planning step | The model reasons about what steps are needed (explicit CoT or implicit) |
| Tool call | The model invokes an external tool: web search, code execution, file read/write, API call |
| Observation | The result of the tool call is returned to the model's context |
| Self-evaluation | The model assesses whether the result is sufficient or further steps are needed |
| Output / handoff | Final answer delivered to user or passed to another system |
The ReAct pattern (Thought → Act → Observe → repeat) is the most studied formalization of this loop.
Types of autonomy
Agentic workflows vary by the degree of human involvement:
- Human-in-the-loop: agent pauses and requests approval before consequential actions.
- Human-on-the-loop: agent runs autonomously but can be interrupted; human monitors.
- Fully autonomous: agent executes end-to-end without human checkpoints (highest risk for irreversible actions).
The appropriate level of autonomy depends on task stakes, reversibility of actions, and reliability of the model's self-correction.
| Term | Distinction |
|---|---|
| AI agent | The system; an agentic workflow is the pattern it follows |
| ReAct | A specific prompt pattern for implementing an agentic workflow |
| Multi-agent orchestration | Multiple agents; an agentic workflow can involve one or many |
| Chain-of-thought | A reasoning technique; CoT may be used inside an agentic workflow but is not itself one |
See also
- Agentic AI vs AI agent
- ReAct
- Multi-agent orchestration
- Tool use vs Function calling
- Agent memory vs Context window
- Agents
