What is an agentic workflow?

What it means at work

Where a single prompt gets one answer, an agentic workflow chains steps: gather inputs, process them, check the result, produce an output, and hand it to a person. It maps well to recurring office tasks such as weekly reports, invoice triage, meeting follow-ups and research briefs. The design question is where the AI acts alone and where a person must approve.

What the sources say

• Anthropic’s engineering guide draws the main line between two designs. Workflows are systems where LLMs and tools are orchestrated through predefined code paths, while agents choose their own steps. Most office “agentic workflows” are the first kind: a fixed sequence with AI inside it.
• The simplest pattern is a chain. Prompt chaining decomposes a task into a sequence of steps, where each LLM call processes the output of the previous one.
• Checks sit between the steps: in a prompt chain, Anthropic says you can add programmatic checks on any intermediate steps to ensure that the process is still on track.
• A chain fits when the job splits cleanly: in its guide, Anthropic calls prompt chaining ideal for situations where the task can be easily and cleanly decomposed into fixed subtasks.
• Start small. The same guide recommends finding the simplest solution possible, and only increasing complexity when needed.
• Research behind it: the ReAct paper explores using LLMs to generate both reasoning traces and task-specific actions in an interleaved manner.
• Tool calls inside steps: the Toolformer paper introduces a model trained to decide which APIs to call, when to call them, what arguments to pass.

An example

Every Friday: pull the week’s numbers from a spreadsheet, draft the summary, compare against last week, flag anomalies, and save a draft email for the manager to review. The trigger and the steps are fixed; the AI handles the reading and writing in the middle.

A common misconception

People assume the goal is to remove humans. A better workflow puts humans where judgment matters and removes the copy-paste around it. A good test: if the step fails silently, would anyone notice? If not, add a check.

Sources

• Anthropic, Building effective agents: “Workflows are systems where LLMs and tools are orchestrated through predefined code paths”
• Anthropic, Building effective agents: “Prompt chaining decomposes a task into a sequence of steps, where each LLM call processes the output of the previous one”
• Anthropic, Building effective agents: “You can add programmatic checks … to ensure that the process is still on track”
• Anthropic, Building effective agents: “ideal for situations where the task can be easily and cleanly decomposed into fixed subtasks”
• Anthropic, Building effective agents: “finding the simplest solution possible, and only increasing complexity when needed”
• arXiv, Yao et al., ReAct: Synergizing Reasoning and Acting in Language Models: “generate both reasoning traces and task-specific actions in an interleaved manner”
• arXiv, Schick et al., Toolformer: Language Models Can Teach Themselves to Use Tools: “a model trained to decide which APIs to call, when to call them, what arguments to pass”