HomeBlog

What Is an Agentic Workflow? What security teams need to understand to govern AI agents

November 16, 2026

•

1 min

What Is an Agentic Workflow? What security teams need to understand to govern AI agents
In This Article

AI is now part of how most work gets done, and many employees use it somewhere across their workday to

  • draft and research with assistants,
  • hand analysis and coding to agents, and
  • connect those agents to the apps and data they use every day.

Cyberhaven Labs’ 2026 AI Adoption Risk Report found that 23% of enterprises have already adopted agent-building platforms, and nearly 50% of developers use coding assistants such as Cursor and GitHub Copilot.

In addition, endpoint-based adoption of agentic AI apps doubled between 2025 and 2026, according to Cyberhaven Labs. Each new assistant or agent changes how a task gets done, who or what touches the data along the way, and where that data ends up.

Cyberhaven calls this way of working an agentic workflow, and protecting these workflows is now a core job for security teams.

This protection starts with a clear definition. AI builders and security teams tend to define agentic workflows in different ways, so let's start with how an agentic workflow is broadly defined.

How Do AI and Security Teams Define Agentic Workflows?

AI builders tend to define an agentic workflow by how the agent works.

Andrew Ng, founder of DeepLearning.AI, described agentic workflows in a 2024 talk at Sequoia Capital as a shift from one-pass answers to an iterative loop, where a model plans, uses tools, reviews its own work, and revises. He expands on the idea in DeepLearning.AI’s Agentic AI course, which organizes agentic workflows into four design patterns:

  • Reflection: The model reviews and improves its own output.
  • Tool use: The model calls search, code execution, APIs, and other tools.
  • Planning: The model breaks a goal into steps and adjusts when a step fails.
  • Multi-agent collaboration: Several agents with different roles work on one task.

At Anthropic, they describe agentic workflows in similar terms. They separate workflows, where models and tools follow “predefined code paths,” from agents, which “dynamically direct their own processes and tool usage,” and group both under the term agentic systems.

And at OpenAI, they have a practical guide to building agents that describes agents as “systems that independently accomplish tasks on your behalf.”

These definitions are useful for building agentic workflows. But what about the data these workflows touch?

A security leader looking at the same workflow wants to know which tools the agents use, but they also want to know

  • what data each agent read,
  • whose access it used,
  • what it produced, and
  • where the result went next, including after a person picked it up.

Cyberhaven builds on these definitions with a broader one. An agentic workflow is any business activity in which an AI app or agent reads, creates, or moves data, from the request that starts it to wherever the result ends up.

Three parts of that definition matter most:

  • It includes the human steps. The person who prompts an assistant and the person who forwards its output are both part of the workflow.
  • It includes every level of autonomy, from a single prompt to agents that hand work to other agents.
  • It covers places where AI runs, including endpoints, browsers, SaaS apps, and cloud services.

Tool calls, planning loops, and agent-to-agent handoffs are where most of the data movement happens, so security teams need visibility into those steps as well as the prompt and the final answer.

This is why Cyberhaven treats the workflow as the unit of protection for AI security.

What Does an Agentic Workflow Look Like in Practice?

Consider a customer success team preparing renewal plans for at-risk accounts. Nothing about the work is unusual, and every step uses a tool the company allows. Figure 1 shows how that work unfolds over one week.

Figure 1: A week of agentic work on a customer success team, from a manager's prompt on Monday to an outreach agent's customer email on Thursday that carries another customer's ticket details outside the company

The first thing to notice is that no single step looks like a breach.

The assistant answered a prompt, the agent on the laptop (the endpoint) used access the manager already had, and the scheduled agent did what it was set up to do. The exposure only shows up when someone traces the data from the source systems, through the analysis and the spreadsheet, and out in an email. That full path is the agentic workflow.

What Are the Parts of an Agentic Workflow?

Every agentic workflow, simple or complex, has the same six parts. Each part maps to a question security and AI leaders need to be able to answer. Figure 2 shows all six, using the renewal example.

  1. Trigger What starts the work: a person’s prompt, an event, a schedule, or another agent. In the renewal example, the trigger moved from a person on Monday to a schedule on Wednesday.
  2. Identity and access Whose credentials the agent uses, and what they reach. Agents usually run with the access of the person or service account behind them, which is often more than the task needs. Giving each agent its own identity, scoped to its task, makes its actions traceable and limits what a mistake can reach. The same goes for the OAuth tokens and MCP server credentials agents use to reach other apps.
  3. Inputs The data the agent reads, including files, SaaS records pulled directly or through MCP servers, pasted text, and earlier AI output.
  4. Actions What the agent does with that data: tool calls, MCP server requests, commands, and plan-and-retry loops. This is where the builder’s definition of an agentic workflow lives. (OpenAI’s guide recommends rating each tool as low, medium, or high risk based on factors such as write access, reversibility, and financial impact.) A read-only tool can still be high risk if the data it reads is sensitive.
  5. Outputs What the agent creates, such as summaries, files, code, and messages.
  6. Handoffs Where the output goes next: another agent, an app, a person, or someone outside the company.
Figure 2: The six parts of an agentic workflow (trigger, identity and access, inputs, actions, outputs, and handoffs) shown with the renewal example

Keep in mind that outputs and handoffs deserve attention. An AI summary of a confidential pricing file is a new piece of data. It may have no file label and match no content pattern, but it still holds the same sensitive information, and it moves like any other data. Insiders already use AI summaries this way to chain AI with older exfiltration channels, and agents can do the same thing by accident. Tracking where AI-generated data goes next matters as much as tracking what went in.

This post focuses on how data moves through normal work. Agents also face deliberate attacks, such as prompt injection and malicious MCP servers, that change what an agent does with the data it can reach. Those attacks raise the same questions about access, inputs, and handoffs.

What Are the Levels of Autonomy in Agentic Workflows?

Agentic workflows differ most in how much the agent decides on its own and whose access it uses. Four levels cover most of what organizations run today, and the renewal example climbed all four in a single week. Figure 3 compares the four levels side by side.

  • Level 1, assisted: A person prompts, and the AI responds. The person is both the trigger and the last handoff.
  • Level 2, delegated: The agent acts on one task a person assigned, using that person’s access. Inputs and actions grow, because the agent decides which tools to call and what data to pull.
  • Level 3, autonomous: An event or schedule starts the agent with no prompt. The identity is often a service account or stored credentials, and nobody reviews each run.
  • Level 4, multi-agent: Agents hand work to other agents. Handoffs multiply, several identities touch the data, and each agent sees only its own step. OpenAI describes two common designs: a manager agent that calls other agents as tools, and a decentralized pattern in which agents hand work to one another.
Figure 3: Four levels of autonomy in agentic workflows (assisted, delegated, autonomous, and multi-agent) with who decides, whose access, a renewal example, and the security question for each

Organizations rarely choose a level on purpose. A workflow starts as a prompt, gets delegated once it proves useful, and gets automated once it proves reliable. Controls built for one level tend to miss the next, so the security question has to move with the workflow.

Does a Simple AI Prompt Count as an Agentic Workflow?

AI leaders often push back on calling a simple chat prompt an agentic workflow. By the builder definitions, which start at the agent, that’s fair. A single prompt involves no planning, no tool use, and no autonomy.

But the data risk starts at level one. Cyberhaven Labs found that 39.7% of data movements into AI tools involve sensitive data, and that 32.3% of ChatGPT usage happens through personal accounts. Those interactions are often the first step of a workflow that later becomes delegated or autonomous.

A definition that starts at level two also splits ownership. AI teams govern the agents, data security teams govern everything else, and the moment a person’s prompt turns into an agent’s first action falls between them. One definition that covers every level gives both teams the same map.

What Should Leaders Ask About Every Agentic Workflow?

Security and AI leaders should be able to answer these questions for any workflow that involves AI, at any level of autonomy:

  1. Which AI apps and agents are in use? Include unsanctioned tools, personal accounts, extensions, and MCP servers alongside approved platforms.
  2. Whose access does each agent use? Know whether it runs as a person, a service account, or another agent, and whether that access fits the task.
  3. What data did it read? Trace each input back to the system it came from.
  4. What did it do, and what did it create? Check whether it changed or deleted data, or reached more data than the task needed. Treat summaries, code, and drafts as new data that keep the sensitivity of their sources.
  5. Where did the output go? Follow it through every handoff, including after a person picks it up and sends it by email, chat, or upload.
  6. Does one policy apply to people and agents? The same rule should hold whether a person or an agent takes the action. OpenAI’s guide recommends that sensitive, irreversible, or high-stakes actions trigger human oversight. Knowing which actions touch sensitive data is what makes that rule enforceable.

Figure 4 returns to the renewal week and marks where each question applies, from Monday’s first prompt to the email that leaves the company.

Figure 4: The renewal workflow mapped across people, endpoint, SaaS and browser, and cloud, with each of the six questions marked at the step where it applies

Define the Workflow Before You Govern the Agent

Most AI governance programs start with an inventory of approved tools. That list shows where AI runs. It doesn’t show which data reaches those tools or where the results go next, and that’s where the risk sits. AI and MCP gateways decide which tools can connect, but they don’t show whether sensitive data moves through them. Starting from the workflow gives security and AI leaders one picture of what agents read, create, and pass along, and that picture holds as a task moves from a single prompt to a fully automated chain.

So the question for leaders is where to start asking the six questions. The answer is level one, with the prompts already happening today, because every workflow that reaches level four started there.

In practice, starting there means four steps:

  • Visibility: Find every AI app and agent in use, including personal accounts, browser extensions, and MCP servers.
  • Identity: Give each agent its own identity, with only the access its task needs.
  • Classification: Treat AI outputs as sensitive whenever their sources are, even when the output carries no label.
  • Control: Apply one data policy to people and agents, and keep it in place as workflows move from prompts to automation.

How Cyberhaven Secures Agentic Workflows

Cyberhaven Flow secures agentic workflows with the same Data Lineage that protects human work.

Figure 5 returns to the renewal week and shows how Cyberhaven protects the data across the week.

Figure 5: How Cyberhaven protects the data across the renewal week, with see-and-protect and classify-and-secure controls at each step before sensitive data leaves the company
  • Lineage: Data Lineage traces data from its origin through copies, summaries, tool calls, and handoffs, so AI output stays connected to its source.
  • Visibility: Agentic Data Security discovers AI apps and agents, records their prompts, responses, MCP tool calls, and file activity, and ties the data they touch to the person or system behind it. It also tells corporate AI accounts from personal ones, so teams can see sensitive data headed to personal tools. Browser extensions extend that view to web-based AI, and connectors add visibility into enterprise AI platforms such as ChatGPT Enterprise and Claude Enterprise.
  • Protection: One data policy applies whether a person or an agent moves sensitive data, and Cyberhaven can warn on or block the data in real time on endpoints and browsers.

AI changed work, and agentic workflows are how that work gets done now. Teams that define them clearly can approve more AI use with confidence instead of defaulting to blanket blocks.

Request a demo to see how Cyberhaven provides Data Security for the Agentic Enterprise across human and agentic workflows.

You can also see how data lineage is the key that unlocks agentic security with “MIT Technology Review: Securing the agentic enterprise with data lineage.”

Frequently Asked Questions About Agentic Workflows

What is an agentic workflow?

An agentic workflow is any business activity in which an AI app or agent reads, creates, or moves data, from the request that starts it to wherever the result ends up. AI builders often define the term more narrowly, as a model that plans, uses tools, reviews its own work, and revises. Cyberhaven’s definition also includes the people, prompts, and handoffs around the agent, because that is where data moves.

What is the difference between an AI agent and an agentic workflow?

An AI agent is the system that does the work: a model with tools and instructions that can act on a person’s behalf. An agentic workflow is the full path of the work, including the person or event that starts it, the data the agent reads, what it creates, and where the result goes next. One workflow can involve several agents and several people.

Does using ChatGPT or another AI assistant count as an agentic workflow?

By most builder definitions, a single prompt isn’t agentic because it involves no planning, tool use, or autonomy. From a data security view, it’s the first level of an agentic workflow. Cyberhaven Labs found that 39.7% of data movements into AI tools involve sensitive data, and many prompts become the starting point for delegated or automated work.

What are the levels of autonomy in agentic workflows?

Cyberhaven describes four levels. At level 1, assisted, a person prompts and the AI responds. At level 2, delegated, an agent completes a task with the person’s access. At level 3, autonomous, an event or schedule runs the agent, often as a service account. At level 4, multi-agent, agents hand work to other agents.

What are the biggest security risks in agentic workflows?

Most of the risk comes from data moving where no one expects it. Agents often run with more access than a task needs, AI outputs such as summaries carry sensitive data without labels, and handoffs between agents, apps, and people spread that data further. A coding agent that reads a file of credentials, for example, has already moved live secrets into an AI tool. Agents can also be targeted by attacks such as prompt injection and malicious MCP servers.

How do security teams secure agentic workflows?

Start with visibility into every AI app and agent in use, including personal accounts and MCP servers. Give each agent its own identity with only the access its task needs. Treat AI outputs as sensitive when their sources are, and apply one data policy to people and agents. Data lineage connects each step, so teams can trace data from its source to the final handoff.

Who should own the security of agentic workflows?

AI teams usually govern the agents they build, and data security teams govern the data. Agentic workflows cross both, so the two teams need a shared definition that covers every level of autonomy, from a single prompt to multi-agent chains. Without it, the moment a person’s prompt turns into an agent’s first action has no clear owner.