An agentic workflow is a process in which a language model is called several times, with tools and checks between the calls, so it can plan, act, review its own work and revise before it hands over a result, instead of writing the final answer in one pass. Andrew Ng set out the idea in a series of letters in The Batch in March and April 2024 and named four design patterns that make a workflow agentic: reflection, tool use, planning and multi-agent collaboration. The term describes how the work around the model is organized, not a particular product, and it covers everything from a simple draft and critique loop to a fully autonomous agent.

What does "agentic workflow" mean?

The cleanest definition is Ng's own, from the second letter in the series: "Instead of having an LLM generate its final output directly, an agentic workflow prompts the LLM multiple times, giving it opportunities to build step by step to higher-quality output."

He explained the problem with a writing analogy. A normal prompt is akin to asking someone to compose an essay from start to finish, typing straight through with no backspacing allowed. An agentic version breaks the same job into steps: plan an outline, decide which web searches are needed, write a draft, read it back to spot weak arguments, then revise. Ng's point was that this iterative process is critical for most human writers, and that models benefit from it in the same way.

So the word to focus on is workflow. What changes is the structure around the model: how many times it is called, what it can look up between calls, and what checks its output has to pass before anything leaves the system.

Where did the term agentic workflows come from?

The four-pattern framing comes from five letters Ng published in The Batch: the overview on March 20, 2024, then one letter each on reflection, tool use, planning and multi-agent collaboration, the last dated April 17, 2024. The first letter opened with a prediction: "I think AI agent workflows will drive massive AI progress this year".

The evidence he offered was a coding benchmark. HumanEval, introduced in a 2021 paper on code generation, measures functional correctness for synthesizing programs from docstrings. Ng's team compared published results from several research groups and reported that GPT-3.5 zero shot was 48.1% correct, GPT-4 zero shot reached 67.0%, and GPT-3.5 wrapped in an agent loop achieved up to 95.1%. Read that as a direction, not a guarantee: it is one benchmark, it is a comparison across different papers, and "up to" is the best result rather than the typical one.

The research underneath was already a year or two old. The ReAct paper of October 2022 had models generate both reasoning traces and task-specific actions in an interleaved manner. Self-Refine, from March 2023, used a single model as the generator, refiner, and feedback provider and reported an average gain of about 20 percentage points across seven tasks. Reflexion, also March 2023, reported 91% pass@1 accuracy on HumanEval, surpassing the previous state-of-the-art GPT-4 that achieves 80%. Ng's contribution was to give these techniques a shared vocabulary that people building products could use.

What are the four agentic design patterns?

Ng defined each pattern in one line in his first letter. The reliability column is his own assessment from the later letters.

Pattern What the model does Ng's view of reliability Business example
Reflection Examines its own work to come up with ways to improve it Works fairly reliably Draft a reply to a supplier, check it against your terms, rewrite
Tool use Calls functions such as web search or code execution to gather information, take action, or process data Works fairly reliably Look up the order in your ERP before answering a delivery question
Planning Comes up with, and executes, a multistep plan to achieve a goal A less mature technology, hard to predict in advance Research a prospect and decide which sources to check next
Multi-agent collaboration Several agents splitting up tasks and discussing and debating ideas Output quality hard to predict A writer agent and a reviewer agent with different instructions

Reflection gets much stronger when the critique is grounded in something outside the model. Ng suggested running its code through a few unit tests or searching the web to double check text, then letting the model reflect on the errors it found. A model criticizing its own prose is useful; a model reading a failed test is far more useful.

Planning is where the behavior starts to surprise you, for better and worse. Ng described a live demo in which his research agent's web search API returned a rate limiting error and the agent pivoted deftly to a Wikipedia search tool he had forgotten he had given it. That flexibility is the point of planning, and it is also why it is harder to test: the agent will do things you did not anticipate.

Is an agentic workflow the same as an AI agent?

Not always, and the difference matters when you scope a project. Ng was explicit that many agentic workflows do not need planning: an agent that reflects on and improves its output a fixed number of times follows a sequence of steps that is fixed and deterministic. Planning only comes in for complex tasks where you cannot specify the decomposition ahead of time.

Anthropic's engineering guide, published on December 19, 2024, uses a different vocabulary for the same territory. It groups all of these designs as agentic systems, then splits them into workflows, where models and tools are orchestrated through predefined code paths, and agents, where the model directs its own process and tool usage. Map the two together and a fixed reflection loop is agentic in Ng's sense but a workflow in Anthropic's. An agentic workflow that uses planning to choose its own next step is an agent in both.

That split is the most important design decision you will make, and we cover how to make it in AI agents vs workflow automation. The short version: the agentic patterns are tools you can add to either shape, and most business processes get the benefit of reflection and tool use without ever handing control of the steps to the model.

What do agentic workflows look like in practice?

The patterns are easiest to see in products you can use today. For a longer list organized by department, see our collection of AI agent examples.

When is an agentic workflow overkill?

When one model call already does the job. Anthropic's advice is to find the simplest solution possible, and only increasing complexity when needed, and it notes that for many applications, optimizing single LLM calls with retrieval and in-context examples is usually enough.

The cost is real and it scales with every loop. Anthropic states that agentic systems often trade latency and cost for better task performance, and in its own data agents typically use about 4× more tokens than chat interactions. Four signals that you do not need the extra loops:

  1. A single prompt with good examples already passes your test cases.
  2. There is nothing external for a reflection step to check against, so the model would only be grading itself.
  3. The task is high volume and low value per item, so a fourfold token bill eats the margin.
  4. You can already write the steps down, which means planning adds unpredictability and nothing else.

How do you build an agentic workflow without losing control?

Add the patterns in the order of their reliability, and measure after each one.

  1. Record a single-pass baseline. Run your real cases through one model call and score the results. This is what every added loop has to beat.
  2. Add reflection with an external check. Validate against a schema, a test, a lookup or a rule, not only against the model's own opinion.
  3. Add the minimum tools. OWASP lists excessive functionality, excessive permissions and excessive autonomy as the root causes of Excessive Agency. Give the workflow only what the task needs.
  4. Cap every loop. Anthropic recommends stopping conditions (such as a maximum number of iterations) to maintain control.
  5. Add planning only where steps cannot be written down, and keep a person in front of anything irreversible: payments, deletions, messages to customers.

If you would rather have this designed and built around your own systems, Qubba builds AI agents and the workflows that contain them, priced with a fixed written quote after a 30 minute call.

Frequently asked questions

Is agentic AI the same as an agentic workflow?

They overlap but are not identical. "Agentic AI" is usually used for systems that act with some independence. "Agentic workflow" describes the process design: multiple model calls, tools and checks.
A workflow can be agentic while every step is still fixed by you, as Ng's own example of a fixed reflection loop shows.

Which agentic design pattern should I try first?

Reflection.
Ng called it relatively quick to implement and said he had seen it lead to surprising performance gains.
Pair it with a real check, such as a unit test or a database lookup, and you get most of the benefit with very little added risk.

Do I need a framework to build agentic workflows?

No.
Ng pointed to AutoGen, Crew AI, and LangGraph as options for multi-agent designs, and Anthropic lists several frameworks too, but its guidance is that the most successful teams use simple, composable patterns and that you should reduce abstraction layers and build with basic components as you move to production.

Are agentic workflows reliable enough for production?

Reflection and tool use are, when they are tested like any other software.
Planning and free-form multi-agent collaboration are less predictable, which is Ng's own assessment, so keep them behind iteration caps, logging and human approval for consequential actions until your test results say otherwise.