Agentic workflows: The future of intelligent automation

Traditionally, automation has meant defining rules and workflows upfront for software to follow. With the rise of LLMs, we can now develop systems that interpret goals, make decisions, and take action across multiple systems, adapting as they go.  This new type of automation is called agentic workflows.

This article explains what agentic workflows are, how they work, when to use them, and what challenges you need to be aware of when using them. It also covers why structured content is so important for the agents working within these systems.

What are agentic workflows?

Agentic workflows involve one or more AI agents that are capable of planning, decision-making, and taking action toward a goal.  Traditional automation, such as robotic process automation (RPA), follows simple predefined steps. By contrast, agentic workflow automation is goal driven. The agent continually takes action, evaluates the results of that action, and decides on the next step, dynamically adapting its behavior based on short- and long-term memory, which includes current context, past interactions, and observed outcomes.

While each AI agent has a certain level of autonomy over its own work, the agentic workflow itself provides structure, acting as a flexible pipeline that coordinates the agents as they work together to reach a well-defined goal.

In practice, the AI agents behind agentic workflows are typically based on large language models (LLMs) that have been configured to have access to different tools and external systems. This allows them to retrieve information, coordinate tasks, and take action across different external systems, rather than just generate text. These external systems need reliable APIs and must return data in a predictable, machine-readable structure so the agent can interpret the results and determine what to do next. 

How do agentic workflows work?

Agent workflow systems usually consist of one or more LLM-powered agents, which handle language and reasoning, and an orchestrator. An orchestrator is software that manages everything else around the agents and their reasoning tasks, including workflows, tool and data access, state management, security, and the enforcement of company policies, and which LLMs to call and when.

Most developers use an agent framework, such as LangChain, CrewAI, AutoGen, or Semantic Kernel, rather than build their own orchestrator from scratch.

Below is a high-level architectural diagram of a simple agentic workflow system.

Where do prompts come from?

The workflow begins when a prompt arrives in the  system. The prompt can come directly from a user or from another system, usually as part of a triggered workflow. For example, you could set up a trigger in Jira to send a prompt to your AI system once a task has been blocked for three days. The prompt could be as simple as "Task X in Jira has been blocked for three days."

Phase 1: Interpret the goal

The orchestrator analyzes the prompt by trying to match it to an existing workflow that already has a defined goal, such as “Reduce the length of time that this task remains blocked, either by helping to resolve it or escalating it appropriately.”

Part of this also includes specifying rules from company policies, such as not spamming people, respecting working hours, avoiding immediate escalation, and notifying the assignee before escalating.

Once the workflow is selected, an agent reasons about what information is needed to achieve the goal and which tools to use to obtain it. In this example, it may need to reason about why the ticket is blocked, who the owner is, how urgent the issue is, and whether any other tasks depend on it. To answer these questions, the agent may call tools such as the Jira, Slack, or Google Calendar APIs.

Phase 2: Planning and execution

Once the goal is clear, the workflow is broken down into smaller, actionable steps and placed in a sensible initial order. This initial order is often designed in advance by developers, but it can change dynamically as the results of each step become clear. In our example, an initial order could be to read the Jira ticket, contact the ticket owner, wait for a reply, and then escalate.

The agent then executes the first step: reading the Jira ticket. To execute a step, the agent may use tools to gather data or trigger real-world actions in another system. Based on what happens during each step, including any responses received, the agent can dynamically adjust the plan by changing the steps or adding new ones. In our Jira ticket example, the agent might read the ticket and notice some comments that suggest the task’s already done but is in a testing phase. Escalating at this stage would come across as heavy-handed, so the agent might decide to delay the escalation step and schedule a follow-up for a later date.

This continuous loop of planning, action, observation, and adaptation is called dynamic orchestration, and it’s one of the key differences between agentic AI workflows and traditional automation.

Phase 3: Reflection and improvement 

Once the workflow has finished, the system (as well as the humans that run it) can reflect on whether the correct goal was chosen, whether the agent had enough information to complete its tasks, and whether the timing of the different steps seems appropriate.

In our Jira example, the system might notice that blocked tasks usually get resolved within 24 hours, so the system might choose to wait longer before escalating.

Components of an agentic workflow

We've seen how an agentic workflow runs end to end. Now let's take a look at the core components that make this possible.

Thinking and reasoning

• Decision-making: Agentic systems can interpret goals, evaluate available information, and select the next action to take, rather than just following a fixed set of steps.

• Planning: Agents can produce explicit, step-by-step task plans.

• Memory: Agents can access short-term memories of what has happened during a specific workflow, which helps them have more context. Long-term memories, such as those built up over multiple workflows, are usually stored in external systems such as databases. 

• Context awareness: An agent can adapt its behavior based on information in its environment, such as conversation history or results from tool calls. 

Action and collaboration

• Tool use: Tools allow the agents to gather data from or take action in other systems, including APIs, databases, MCP tools, search engines, and external services. This allows agents to take action beyond simply generating text.

• Dynamic orchestration: This refers to the ability to change the workflow based on new information — for example, changing what to do next based on what just happened. It's where the orchestrator routes and executes steps based on the agent’s earlier decisions.

• Multi-agent systems: In more complex agentic systems, there can be multiple agents, each with specialized roles, and an orchestrator coordinating them. For example, one agent could specialize in retrieving information, another in reasoning over it, and another in communicating results to users.

Benefits of agentic workflows

Agentic workflows can make your organization even more efficient by automatically handling complex tasks involving a combination of reasoning, decision making, and action, such as monitoring queues and then chasing people for updates. This doesn't replace traditional automation; if a task is simple and straightforward enough, it’s much easier to set up a task using something like RPA. However, for more complex tasks, AI agentic workflows allow you to automate tasks that would have previously needed human involvement.

Why agentic workflows need structured content

For agentic workflows to work reliably, they need structured, machine-readable data. When agents take action across different systems, they use APIs that expect content in a predictable format, such as JSON. If the content is unstructured — for example, presented as a single long page of text — it’s hard for the agent to extract the useful parts.

It's much easier for agentic systems to work with content that’s broken down into small, well-defined pieces. Platforms like Contentful already help teams structure their content, including e-commerce listings and help articles, and provide validation and governance so the same wording and information can be used consistently across different channels and workflows.

When to use an agentic workflow

It often becomes clear early on when a problem is too complex for traditional, rules-based automation.  But there are a number of other AI-related options to consider before jumping straight to an agentic workflow.

In many cases, a simple LLM, or an LLM with tool access, will suffice.  This works well when the sequence of tasks is fixed and won't need to change during the workflow. For example: search for some information, summarize it, format it, and send it to someone. At no point does the order need to change.

Agentic workflows become more useful when the system needs to dynamically change steps as it progresses, choosing what to do next based on what has happened previously. 

If you choose an agentic workflow, it is worth considering what happens if the agent makes an incorrect decision. It’s often a good idea to keep a human in the loop or add human approvals at key stages in the workflow.

Challenges of agentic workflows

Reliability is a key challenge that agentic workflows face: agents can hallucinate information, struggle with ambiguous instructions, or get stuck in unhelpful loops. It’s important to have robust error handling, clear guardrails, and fallback options in place. 

Prompt engineering and behavior design are also important to get right. Because LLM agents are non-deterministic, the same prompt can lead to different action paths, which makes testing and observability harder. Additionally, the different ways prompts, guardrails, and policies are expressed can affect how agents perform. Allowing time to iterate and test this is important.

Deciding how much autonomy to give your agents is also a big challenge — the more freedom you give agents, the more useful they can potentially be to you. However, you don't want to give your agents more privileges than they need (such as giving an agent permission to delete records when it only needs read access), as this leads to an increased risk that it’ll behave in unpredictable and problematic ways. 

Finally, practical considerations such as latency, cost, and security still apply, especially when agents are acting across live systems. Agents may also require memory management, which includes deciding what context to retain and where to store it.

Contentful as the content provider for agentic workflows

Agents need to work with structured content, and Contentful provides this through predictable APIs. Content is organized into a clear content model of your design, and validation and governance is added so you can maintain style and brand consistency while also following organizational policies.

Agents can use Contentful's Content Delivery and Content Management APIs to retrieve content, analyze or audit it, and then update it or publish changes. This means Contentful provides the source of truth for agents to get and update content.

Take a tour

Chat with our team

Your experiences should wow, not wait. Let's talk.