Building with AI

Agentic workflows

4 min readintermediateUpdated 28 Sept 2026
1 · In one line

An agentic workflow runs a task as several language model and tool steps, joined by a path that your code fixes in advance.

1 · What it is

People use the phrase loosely. OpenAI’s guide calls any series of steps that meets a user’s goal a workflow. It counts a system as an agent when a model runs those steps and makes the decisions. Anthropic puts every version under one label, agentic systems. It then splits them in two. In workflows, language models and tools follow paths written in code ahead of time. In agents, the model directs its own process and picks its own tools. This page uses the narrower, workflow sense.

Anthropic describes several common shapes. Prompt chaining is the one in the figure. Routing sorts each incoming request into a type. It then passes the request to a prompt built for that type. That way easy questions can go to a smaller, cheaper model and hard ones to a stronger model. Parallelization means running parts at the same time, and Anthropic describes two kinds. Sectioning splits a job into parts that do not depend on each other and runs them side by side. Voting asks the same question several times and compares the answers. In orchestrator-workers, one lead model splits the job into pieces. It gives each piece to a helper model, then merges the answers. The lead model only decides the pieces after it sees the request, so, unlike parallelization, its pieces are not fixed in advance. In evaluator-optimizer, a writer call drafts an answer and a critic call grades it and suggests fixes. The pair repeats. It works like a student and a teacher passing an essay back and forth.

A related research method, Self-Refine, lets one model do every job: it writes a draft, gives feedback on it, then rewrites it. Averaged over 7 tasks, from chat replies to maths problems, it scored roughly 20 percentage points higher than answering in one go.

Frameworks are ready-made code libraries, and they turn these shapes into building blocks. Google’s Agent Development Kit (ADK) has template workflow agents. They run smaller sub-agents in order, side by side or in a loop, without asking a model how to arrange them. The loop version keeps cycling through its sub-agents and stops only when a set exit rule is met. In the Python and Go versions of ADK 2.0, graph-based and dynamic workflows have replaced these templates. LangGraph adds shared features, such as persistence (saving the state of a run), to both workflows and agents.

2 · Why it exists

One big prompt is hard to control, and a free-running agent is hard to predict.

One call, many jobsWhen a task has several parts to weigh, models generally do better when each part gets its own call.
Hard to see insideWhen 20 people tested it, splitting work into chained model steps improved results and made the system more transparent and controllable.
Agents cost moreA model that steers itself brings higher costs and the risk that one mistake feeds the next.
3 · How it works

Follow one joke through a three-call chain with a code gate.

The model writes the jokes. A plain code check decides which call runs next.
  1. 1 · splitA developer breaks the job into a fixed sequence of steps, where each model call takes the result of the step before it.
  2. 2 · generateIn the LangGraph example, the first model call writes a short joke about a topic.
  3. 3 · gateA plain code check, not a model, tests whether the joke contains a question mark or an exclamation mark.
  4. 4 · branchA pass ends the run, while a fail sends the joke to a second call that adds wordplay and a third that adds a surprising twist.

In a workflow your code picks the next step. In an agent, the model does.

4 · Where it's used
WhoWhat they askWhat it works with
Marketing team“Write the launch post, then translate it into Spanish.”A two-call chain, draft first and translation second
Support lead“Send refund requests and technical questions to different handlers.”A routing step that picks one downstream prompt
Security reviewer“Check this code for vulnerabilities in several ways.”Parallel prompts that each flag problems they find
Translator“Keep refining until the meaning matches the original.”A writer call and an evaluator call in a loop
5 · What it solves, and what it doesn't
solves
  • For well-defined tasks, a fixed path gives predictability and consistency.
  • Each model call gets an easier job, trading extra waiting time for higher accuracy.
  • Code checks between steps can confirm that a run is still on track.
  • Because code sets the route, you can forecast how long a run takes, what it costs and how well it does.
doesn't solve
  • Open-ended problems where the number of steps cannot be predicted need an agent instead.
  • Plenty of apps never need a chain, since a single carefully prompted call, given retrieved documents and a few examples, does the job.
  • Frameworks can hide the underlying prompts and responses, which makes bugs harder to trace.
  • Graphs that declare every branch up front can get cumbersome as a workflow grows more dynamic.
6 · Go deeper

Sources used

This explainer is written in original language. The links below support its factual claims.

  1. officialBuilding effective agents, Anthropic · read 28 Sept 2026
  2. docsWorkflows and agents, LangChain · read 28 Sept 2026
  3. docsAgent orchestration, OpenAI · read 28 Sept 2026
  4. officialA practical guide to building agents, OpenAI · read 28 Sept 2026
  5. docsTemplate agent workflows, Google Agent Development Kit · read 28 Sept 2026
  6. paperAI Chains: Transparent and Controllable Human-AI Interaction by Chaining Large Language Model Prompts, Wu et al., CHI 2022 · read 28 Sept 2026
  7. paperSelf-Refine: Iterative Refinement with Self-Feedback, Madaan et al. · read 28 Sept 2026