Agentic Workflow

An agentic workflow is a multi-step process in which an AI model is called repeatedly, using tools and feedback, to build toward a result instead of producing a final answer in one pass. The steps may follow a path the developer defines in code, or the model may choose them as it goes. Either way, the work is broken down, checked and refined along the way.

How an Agentic Workflow Works

Andrew Ng popularized the term in March 2024, describing an agentic workflow as one that prompts an LLM multiple times, giving it opportunities to build step by step toward higher-quality output. He grouped the techniques into four design patterns:

• Reflection: the model critiques its own output and revises it.
• Tool use: the model calls external functions, such as search, code execution or a database query, through tool calling.
• Planning: the model breaks a goal into steps before carrying them out.
• Multi-agent collaboration: several specialized model instances, each with its own instructions and tools, work on different parts of a problem.

Anthropic's guidance describes common structures for workflows whose path is set in code:

• Prompt chaining: each call processes the output of the previous one.
• Routing: an input is classified and sent to a specialized path.
• Parallelization: independent subtasks run at the same time, or the same task runs several times and the results are compared.
• Orchestrator-workers: a central model splits the task and delegates parts to worker calls.
• Evaluator-optimizer: one call produces a draft and another evaluates it in a loop until it meets the bar.

Why Agentic Workflows Matter

Single-pass generation has no chance to notice its own mistakes. Breaking a task into steps gives the system points where it can check, correct and use real data, which is why multi-step setups often outperform one long prompt on complex work.

Structure also brings predictability. When the path is fixed in code, each step can be tested on its own, failures are easier to locate and costs are easier to estimate. The trade-off is more model calls, more latency and more to maintain, so the usual advice is to add steps only where they measurably improve the result.

Agentic Workflow Example

A team drafts sections of a product requirements document from customer interview notes. Step one extracts candidate requirements from the notes. Step two drafts each requirement. Step three is an evaluator call that checks every requirement has acceptance criteria and a source quote, and flags any that do not. Step four revises the flagged items. Step five sends the draft to a product manager for review. The order never changes, so in Anthropic's terms this is a workflow; in Ng's broader sense, it is an agentic workflow built from reflection and tool use.

Agentic Workflow vs. AI Agent

The terms overlap, and sources use them differently. Anthropic separates workflows, where LLMs and tools follow predefined code paths, from agents, where the model directs its own process and tool use. Ng's "agentic workflow" covers both, as long as the model is called iteratively instead of once. In practice, if your code decides the next step, you have a workflow. If the model decides, you have an AI agent. Many production systems combine both: a fixed outer workflow with an agent handling one open-ended step.