An agentic workflow is a multi-step process in which an AI agent decides some of the next steps while it runs, instead of following a path you scripted in advance. You set the goal and the limits, the agent chooses its route at the points where the route depends on what it finds, and fixed steps handle everything else.

That mix is the point. An agentic workflow sits one layer above a single AI agent, which handles one task, and it's how most companies actually put agentic AI to work: a mostly predictable process with a few places where software gets to use judgment.

Most things sold under the name don't have those places. They're linear automations with a language model dropped into one step, running the same path on every pass. That can be a perfectly good build, but it isn't agentic. The opposite mistake, making every step agentic, is the more expensive one.

TL;DR

• A workflow is agentic when at least one step picks the next move from what the previous steps found. A fixed path with a model writing inside one step is an AI workflow, which is often the better build.

• Mark every step as fixed or agentic. In a typical inbound flow, identifying the visitor, routing, sending and logging stay fixed, and only judging intent, researching an angle and drafting get an agent.

• Andrew Ng's four patterns (planning, tool use, reflection and multi-agent collaboration) cover almost every build. In the coding results he published, GPT-3.5 went from 48% to as high as 95% once it ran inside an agent loop, ahead of GPT-4 prompted once.

• Pick the smallest shape that fits: a single agent in a loop, an orchestrator with workers, or a writer paired with a reviewer. Every extra agent adds a handoff where errors can hide.

• Each decision point costs time, tokens and a chance of a wrong call, and the odds multiply. Five steps that are each right 90% of the time make a chain that's right about 59% of the time.

• Build the fixed skeleton first, give each agentic step a fallback to a person, require sign-off on everything a buyer will read, and log every decision so a bad run can be replayed.

Agentic workflow or automation? Find the step that chooses

Every workflow has moments where something decides what happens next. In plain automation, you made all of those decisions when you built it. A Zapier zap or an n8n flow that runs A, then B, then C will run A, B and C again tomorrow whatever the data says, and any branch it takes is a branch you wrote.

In an agentic workflow, at least one step makes that call at runtime. It looks at the goal and at what earlier steps turned up, then picks the next action: read the company's careers page, check for recent funding, or skip both because the website already answered the question.

An AI step on its own doesn't change that. A model that writes a personalized first line inside a fixed sequence is doing work, not choosing a route, and that's an AI workflow. Neo4j's guide to agentic workflows boils the choice down to one line: if the path is known, use a deterministic workflow, and if the goal is clear but the path isn't, consider an agentic one.

Mapping an inbound flow: which steps get to decide

The hybrid is easiest to grasp on a real process with every step labeled. Here's an inbound flow for a known company landing on your pricing page, the kind of thing we build in our website visitor workflows.

Three steps out of seven get agency, and the agent never touches routing, sending or the CRM write-back. That's the shape that holds up in production. The plumbing that catches the signal and tells the rep stays predictable, and the agent only handles the judgment in the middle, which is also where a fast, relevant answer matters most.

Outbound splits the same way. Moving a record, sending a templated email or updating a field is fixed work that plain sales automation does more cheaply. Researching an unfamiliar account is different, because the right sources and the right angle change from company to company. We've covered which prospecting tasks to automate with AI and where a person should stay in charge.

This is also why fully autonomous outbound so often disappoints. Teams make the whole motion agentic, including the steps that should have stayed fixed, and get something slow, costly and hard to predict.

Ng's four patterns, and what each one buys

In March 2024, Andrew Ng described four design patterns for agentic workflows, and nearly every build since uses some of them. His case rested on a coding benchmark. In the first letter in that series, he shared results compiled from several research teams on HumanEval: GPT-3.5 prompted once got 48.1% right and GPT-4 got 67.0%, while GPT-3.5 wrapped in an agent loop reached as high as 95.1%. The loop around the model mattered more than the model upgrade.

Planning

The agent breaks a goal into steps and decides their order. You get flexibility on jobs that vary, and you give up some predictability, because two similar inputs can take different routes.

Tool use

The agent calls APIs, searches the web and queries your CRM, so it can act instead of describe. Its ceiling is set by the tools you hand it, and every tool is also a permission you'll want to think about.

Reflection

The agent checks its own output against the brief and improves it before moving on, the way a person rereads a draft. Quality goes up, and so do time and tokens, since every pass is another call.

Multi-agent collaboration

Several agents take different roles, say one researching and one writing, and pass work between them. That widens what the workflow can take on, at the price of more handoffs to watch.

You don't need all four. Planning plus tool use already makes a step agentic, and most GTM builds stop there.

Pick the smallest shape that holds the job

The patterns combine into a few recurring shapes. Anthropic's engineering team describes the common ones in its guide to building effective agents, and they map neatly onto sales work.

A single agent in a loop plans, acts, checks and repeats until it's done. That covers most account research and drafting, and it's where to start.

In the orchestrator-workers shape, a central model breaks a task into pieces at runtime, hands them to worker models and combines what comes back. Anthropic suggests it for tasks where you can't predict the subtasks ahead of time. In GTM, picture researching a list of 40 accounts where each one needs different lookups. The weak point is the first split. If the orchestrator divides the job badly, every worker does the wrong thing well.

The writer-and-reviewer shape (Anthropic calls it evaluator-optimizer) has one model draft and another score the draft against clear criteria and send it back. It catches a lot of weak output and roughly doubles the calls, so save it for work where quality decides the result, like a first email to a target account.

A crew of researcher, writer and reviewer looks impressive, but every handoff is a place for an error to hide. If the researcher pulls a wrong fact, the writer builds on it and the reviewer may only judge the tone, so the mistake reaches the output looking polished. Most GTM jobs are better served by one well-scoped agent with a lot of fixed plumbing around it.

Why every decision point has a price

Agency isn't a free upgrade. Each step where the agent decides adds waiting time, spends tokens and carries a chance of a wrong call. A fixed step behaves the same way every run, and when it fails, the log points straight at the step that failed. An agentic step can surprise you, and working out why means piecing together what it saw and what it chose.

The odds multiply along a chain. If each agentic step is right 90% of the time, five in a row are right only about 59% of the time (0.9 to the fifth power). Anthropic's guide makes the same point in plainer terms, warning that more autonomy brings higher costs and a risk of errors compounding. That's why long, fully agentic chains dazzle in a demo and turn flaky in production.

Cost creeps as well. A workflow that picks up a few extra agentic steps over a year can end up slower and pricier than the plain automation it replaced, and nobody notices until the invoice does. Keep agentic steps few and fixed steps many. If a person could write the rule in advance, write the rule.

Building one: skeleton first, agency second

1. Draw the fixed steps. Lay out the whole process as if no agent existed, then mark the one or two places where the route changes from case to case. Only those become agentic.

2. Test each candidate. A step should get an agent only if you couldn't write its rule in advance and its path really varies between runs. What a wrong answer would cost decides whether a person checks it.

3. Give every agentic step a way out. Define what happens when the agent is unsure or stuck, usually a handoff to a person along with what it found so far. Without one, a confused agent stalls or improvises.

4. Gate what buyers see. Anything that reaches a prospect waits for a human in the loop to approve it. n8n lets you require human approval before an AI agent uses a specific tool, so the agent can research freely but has to wait before it sends a message or changes a record. One of our website visitor workflows builds the same approval step into the flow.

5. Log every decision. Keep the workflow under the workflow orchestration that runs the rest of your GTM system, and record what the agent looked at and why it chose each step, so a bad run can be replayed instead of guessed at.

Clean data matters at every step. An agent reasoning over stale job titles or duplicate companies makes the same bad call a person would, only faster and more often.

Clay is where many of these workflows get assembled, because it sits between the data and the action, and an agentic step can run there as a research column. Putting a research agent where a simple lookup would do is one of the most common reasons we see Clay builds get expensive, and we've written more broadly about why most Clay implementations fail.

In short

An agentic workflow is a mostly ordinary process with a few steps that get to think. Its value depends on choosing those steps well: the ones where the route genuinely differs between runs, and where a wrong turn is cheap or gets caught before a buyer sees it.

Everything else should stay fixed, because fixed steps are fast, cheap and easy to debug. The builds that last look a bit dull on paper, with a sturdy skeleton, one or two decision points, a fallback to a person, an approval step and a log someone reads. If the result isn't clearly better than the simpler version, pull the agency back.

When you want help drawing that line, we build agentic workflows on top of the Clay, n8n and CRM setup you have now, write down which steps decide and why, and hand the finished build to your team to run.

Frequently asked questions