Ten thousand words long article! From AI Agent to Agent Workflow, a detailed understanding of Agentic Workflows

introduction

Concepts such as AI Agent, Agentic AI, Agent Architecture and Agent Workflow are gaining a lot of attention nowadays, but what exactly are they? What can they do? New technologies are often accompanied by confusing terms and hype. This article will deeply analyze the key concept of agent-based AI - Agentic Workflows .

AI agents themselves do not have much practical use. Only by giving them roles, goals, and structures, and only through workflows, can they really play a role . Understanding agent workflows helps us understand the operation and logic of AI agents. To this end, this article will take you from AI Agents to Agent Workflows, and take you to a detailed understanding of Agentic Workflows. The specific content is arranged as follows:

• What is an AI Agent?
• What is Agent Workflow?
• What does Agent workflow consist of?
• Agent Workflow vs Agent Architecture
• Agent workflow mode
• Agent workflow application scenarios
• Advantages and disadvantages of Agent workflow

What is AI Agent?

An AI agent is a system that combines the reasoning and decision-making capabilities of large language models (LLMs) with real-world interaction tools to enable it to complete complex tasks with limited human intervention. Agents are given specific roles and have varying degrees of autonomy to achieve their end goals. They also have memory capabilities , are able to learn from past experiences, and improve their performance over time.

To better understand the role of AI agents in agent workflows, it is necessary to first understand their core components.

Components of an AI Agent

Although AI Agents are designed for semi-autonomous decision making , they still rely on a larger framework of components to function properly. These components include:

• LLMs (Large Language Models) : Enabling agents to reason
• Tools : Help agents complete tasks
• Memory : enables agents to learn from past experiences and optimize responses over time

Reasoning

The power of AI agents lies in their iterative reasoning ability , that is, continuous "thinking" throughout the problem-solving process. The reasoning ability mainly comes from the underlying LLM and plays two core roles: planning and reflecting .

• Planning : The agent will decompose tasks (Task Decomposition), breaking down complex problems into smaller, more executable steps. This allows the agent to perform tasks systematically and choose different tools based on different needs. In addition, the agent can also perform query decomposition (Query Decomposition), which is to break down complex queries into simpler sub-queries to improve the accuracy and reliability of LLM responses.
• Reflecting : The agent reviews the results of its actions and makes adjustments based on the results and external data to optimize subsequent decisions.

Tools

Since LLM is limited to the knowledge acquired during its training ( static parameterized knowledge ), to extend its capabilities, AI agents can use external tools such as:

When LLM selects the right tool to complete a task, it performs a function call , thereby expanding its capabilities beyond simple text generation to enable interaction with the real world.

The choice of tool can be pre-defined by the user or dynamically determined by the agent . Dynamic tool selection helps solve complex tasks, but for simple workflows, pre-defined tools may be more efficient.

Memory

The memory capability of AI agents is one of the key features that distinguishes agent workflows from pure LLM workflows. Memory allows agents to store context and feedback across multiple user interactions and sessions , enabling a more personalized experience and optimizing long-term performance .

The memory of an AI agent can be divided into two types:

• Short-term Memory : Stores recent interaction information, such as conversation history, to help the agent decide the next action.
• Long-term Memory : Stores information accumulated over a long period of time and learns across sessions to achieve personalization and continuous optimization.

What is Agent Workflow?

In general, a workflow  is a series of interrelated steps designed to accomplish a specific task or goal. The simplest workflows are deterministic , meaning they follow a predefined sequence of steps and cannot adapt to new information or changing circumstances. For example, an automated expense review workflow might look like this: If the expense is tagged "dining" and the amount is less than $30, then automatically approve it.

However, some workflows leverage large language models (LLMs) or other machine learning models to enhance their capabilities. These are often referred to as  AI workflows and can be divided into two types: agentic and non-agentic :

• Non-agent AI workflow : LLM generates output based on input instructions. For example, the process of a text summarization workflow may be: receive a long text → let LLM summarize → output summary. This process only relies on the text processing ability of LLM and does not have the ability to make autonomous decisions or perform tasks, so it does not belong to the agent workflow.

• Agent-based AI workflow : A series of steps dynamically executed by one or more AI agents to complete a specific task. Agents have a certain degree of autonomy within the scope of permissions granted by users , and can collect data, perform tasks, and make actual decisions. In addition, agent-based workflows use the reasoning ability, tool use ability, and persistent memory of AI agents to make traditional workflows more responsive, adaptive, and self-evolving .

Agent workflow composition

For an AI workflow to become an agent-based workflow , it must have at least the following three core characteristics:

1. Make a plan:  Agent-based workflows start with planning . LLM is responsible for task decomposition , breaking down complex tasks into smaller subtasks and determining the best execution path.
2. Execute actions with tools  Agent-based workflows use a set of predefined tools (such as APIs, databases, search engines, etc.) and corresponding permission management to execute tasks and implement planned plans.
3. Reflect and iterate   agents can evaluate results at each step, adjust the plan if necessary, and iterate until satisfactory results are achieved.

Three workflows

We can distinguish three different types of workflows:

The differences are mainly reflected in:

• AI workflow vs. traditional workflow : AI workflow relies on AI models, while traditional workflow relies on predefined steps.
• Agent-based vs. non-agent AI workflows : Agent-based AI workflows use dynamic AI agents , while non-agent AI workflows simply use static AI models to process tasks.

Therefore, agent-based workflows are more adaptive and dynamic than non-agent AI workflows.

The difference between Agent architecture and Agent workflow

As new technologies evolve, so does the terminology. Although some people use the terms “Agentic Architectures” and “Agentic Workflows” interchangeably , there is a clear distinction between the two.

• Agent workflow  focuses on the process of task execution , that is, how agents break down tasks, use tools, adjust strategies, etc.
• Agent architecture  focuses on system design , that is, how to build an overall framework that includes agents, tools, and memory systems.

Agent workflow mode

To recap, an agentic workflow refers to a structured series of steps that an agent takes to accomplish a specific task (end goal). So when we talk about agentic workflows, we are actually talking about a specific behavior pattern of an agent to achieve its end goal .

The AI ​​agent core components mentioned above play a key role in the agent workflow model:

• Reasoning ability  promotes the realization of planning pattern  and reflection pattern  ;
• Tool use capability  is  the basis of tool use pattern .

1. Planning Pattern

Planning mode  enables agents to autonomously break down complex tasks into a series of smaller, simpler subtasks, a process called task decomposition. Task decomposition can improve the quality of workflows . Specifically, it can: reduce the cognitive load of LLMs and reduce their computational pressure; improve reasoning ability and enhance problem-solving ability; reduce hallucinations and other inaccuracies and improve the reliability of outputs. Planning mode is particularly effective when the path to the goal is unclear and requires flexible adaptation to the problem-solving process . For example: When an AI agent is instructed to fix a software bug  , it may decompose the task as follows : read the bug report, identify relevant code snippets, generate a list of possible error causes, select an appropriate debugging strategy, run the repair code and observe the error message, and adjust the plan if it fails.

While planning mode helps agents better solve complex tasks, it can lead to less predictable results than deterministic workflows. Therefore, planning mode is suitable for tasks that require deep reasoning and multi-step reasoning .

2. Tool Use Pattern

Generative LLMs have a significant limitation: they can only answer based on existing training data and cannot retrieve information or verify facts in real time . Therefore, LLMs may: generate false information ("hallucinations") and "guess" answers when faced with uncertain questions. This problem can be solved by RAG, which enhances the accuracy of LLM responses by providing relevant, real-time external data , but it is limited to data retrieval rather than dynamic interaction with the external environment. The tool usage pattern  further expands the agent's capabilities, enabling it to dynamically interact with external systems rather than just passively retrieve data. The tool usage pattern is suitable for tasks that require real-time data support or rely on external resources , such as: searching for the latest news on the web; obtaining user history from a database; and automatically sending emails. The following are some common tools and their uses.

3. Reflection Pattern

Reflective mode is a self-feedback mechanism where the agent iteratively evaluates the quality of its output or decision  before taking the final action, and optimizes its reasoning process accordingly. This process can help the agent: correct errors and reduce inaccuracies; continuously improve and improve the quality of decisions; enhance adaptability and better meet user needs. Reflective mode is particularly useful for tasks that are difficult to perform successfully in one go , such as code generation :

1. Agent generates code snippets ;
2. Run code in a sandbox or execution environment ;
3. Get error information and feed it back to LLM;
4. Let LLM  iteratively optimize the code until it executes successfully.

The power of reflection lies in the agent's ability to critique its own output and dynamically incorporate these insights into the workflow, enabling continuous improvement without direct human feedback. These reflections can be encoded in the agent's memory, allowing for more efficient problem solving during the current user session, as well as personalization and improved future interactions by adapting to user preferences.

Agent workflow application scenarios

Atomic Design Patterns , such as Planning and Tool Use, can be combined in many ways to effectively use agent AI to handle various tasks in different fields. In addition to combining design patterns, AI agents also have different tool combinations and even dynamically select tools to adapt to task requirements, enhance their understanding of task goals by incorporating human feedback loops , and grant them different levels of autonomy and decision-making power .

These diverse configurations enable agentic workflows to adapt to a wide range of industry needs. The following are two representative application cases: Agentic RAG , Agentic Research Assistants , and Agentic Coding Assistants .

1. Agentic RAG

Retrieval-augmented generation (RAG)  is a framework that enhances LLM generation capabilities through external data. Agent-RAG introduces one or more AI agents into the RAG process to make it more intelligent and dynamic. Specifically:

In the Planning Phase : The agent can break down complex queries into smaller sub-queries (Query Decomposition) and determine whether it needs to request additional information from the user to complete the task more accurately.

In the data retrieval and evaluation phase : the agent can evaluate the relevance and accuracy of the retrieved data to avoid providing useless or erroneous information. When the query results are not satisfactory, the agent can readjust the query , return to the query decomposition phase, or even formulate a new query plan .

2. Proxy Research Assistant

Agentic Research Assistants , sometimes referred to as "Deep Research" by AI companies , are used to generate in-depth reports and detailed insights into complex topics . They are based on agent RAGs , but instead of just retrieving information, they can analyze and synthesize data to provide more in-depth results. Here's a comparison between agent research assistants and traditional RAGs:

Specifically: Proxy research assistants typically use LLMs that are fine-tuned for web browsing, task decomposition, and dynamic programming ; they are able to proactively request additional information from users to more clearly understand the task objectives; and they adjust the research direction based on the retrieved information , explore new angles, and ensure that complete data is obtained.

The agent research assistant not only retrieves information, but also identifies trends and forms systematic insights. In addition, it can analyze data across time periods and write detailed research reports, improve the efficiency of researchers, and reduce the time for manual search and analysis. Currently, OpenAI, Perplexity and Google  have launched their own deep research products.

3. Proxy Coding Assistant

Agentic Coding Assistants  can complete code generation, refactoring, optimization, and debugging with minimal human intervention , support creating PRs and submitting code, and improve team collaboration efficiency. In contrast, non-agent coding assistants (such as early versions of GitHub Copilot) are limited to code generation and lack environmental interaction and adaptive capabilities.

The core features of the proxy coding assistant are as follows:

1. Execution and self-improvement : Not only can code be generated , but it can also execute the code and iteratively optimize it based on error information.
2. Code repository management permissions :

• Agents can create commits and pull requests to automate software development processes (such as Anthropic Claude Code ).
• Agents can ask for user confirmation before executing , ensuring that humans are in control of key decisions. (e.g. Cursor's Agent ).

• Long-term Memory :
• The agent is able to remember its mistakes and improve itself , performing more intelligently on future tasks.

Agent workflow practical application case

After introducing the application scenarios of agent workflows, let’s explore the workflows of two actual agent systems: Claygent (Clay) and  ServiceNow AI Agents . Each agent workflow uses a unique combination of patterns and tools, gives agents different degrees of autonomy and decision-making capabilities, and relies on different levels of human feedback and participation.

Claygent is an AI research agent launched by a data enrichment and outreach automation company, which mainly serves growth teams and sales teams to help them solve the tedious tasks of prospect research and data enrichment . Its workflow is to enrich LinkedIn profiles and send personalized introduction messages as an example:

• The user enters a list of names and emails and specifies the required data fields (eg, work experience, education, skills).
• The agent formats the query using preconfigured prompt templates, processes the query through LLM and utilizes web crawlers to search for relevant URLs on LinkedIn and extract the required data from the profiles.
• The agent can then call another LLM to summarize or analyze the acquired data and generate personalized outreach messages to suit each prospect.

Claygent has a flexible and customizable workflow, allowing users to creatively customize tasks; pre-configured prompt templates ensure that agents follow specific task objectives and improve the quality of results; and multi-agent collaboration, with different LLMs handling different tasks (data capture, analysis, personalized messages).

ServiceNow AI Agents is a new feature of the ServiceNow cloud platform, which focuses on automating repetitive tasks in IT, operations, HR and customer service, optimizing existing workflows, and ensuring that the final decision-making power is in the hands of humans . Take the processing of technical support tickets as an example:

• After the user submits a work order, the agent workflow is triggered. The agent calls RAG (Retrieval Enhanced Generation) technology to search for relevant information in the internal IT support knowledge base, summarize the search results, and analyze similar cases.
• The agent then creates a summary that the IT specialist reads, covering the problem and possible solutions, and generates a recommended action that the IT support specialist approves or rejects.

ServiceNow AI Agents features human-machine collaboration, where agents only assist IT experts and do not directly execute decisions; they are safe and controllable, where agents can only perform tasks within a limited scope to avoid affecting the end-user or customer experience; and they feature structured workflows, where agents strictly follow predefined rules to reduce unexpected behaviors.

Advantages and disadvantages of Agent workflow

Agentic AI is rapidly moving from the machine learning community to the mainstream. Given the excitement, anticipation, and expectations surrounding agentic AI, it can be difficult to sort the gap between hype and reality and understand its true capabilities and limitations. In this section, you will be provided with a comprehensive perspective on the strengths, challenges, and limitations of Agentic workflows.

Advantages of Agent Workflow

Agentic workflows go beyond traditional automation by enabling AI agents to plan, adapt, and improve over time. Unlike deterministic workflows that follow fixed rules, agentic workflows are able to dynamically respond to complexity, fine-tune their approach through feedback, and scale to handle more complex tasks. This adaptability makes them particularly important in scenarios that require flexibility, learning, and decision-making.

Advantages of Agentic Workflow:

• Flexibility, adaptability, and customizability . Static, deterministic workflows have difficulty adapting to changing situations and unexpected difficulties. Agentic workflows, on the other hand, offer the flexibility to adjust and evolve based on the difficulty of the task, ensuring they always remain relevant and provide the best solution. They are also customizable by combining different modes, providing a modular design that enables iterative upgrades as needs and complexity increase.

• Improved performance on complex tasks . By breaking down complex tasks into smaller manageable steps (via task decomposition and planning), Agentic workflows significantly outperform deterministic, zero-shot approaches when handling complex tasks.

• Self-correcting and continuous learning . Reflection mode allows Agentic workflows to evaluate their own behavior, improve strategies, and improve results over time. Using short-term and long-term memory, they learn from past experiences, making each iteration more efficient and personalized.

• Operational efficiency and scalability . Agentic workflows can automate repetitive tasks with high accuracy (if designed properly), reducing manual work and operational costs in specific scenarios. They can also be easily scaled to handle larger workloads or complex systems.

It’s important to note that AI agents are still an emerging technology, and the above list of benefits is likely to expand as researchers and users discover new ways to incorporate agents into their workflows.

Agent Workflow Limitations

Despite their many benefits and innovative capabilities, AI agents also present some challenges and limitations. Due to their probabilistic nature, AI agents inherently add complexity to workflows. Just because agents can be used to automate processes, it doesn’t mean they should be used. Here are some notable challenges and limitations of agentic workflows:

• Excessive complexity for simple tasks . AI agents can add additional overhead when used for simple workflows, such as form entry or basic data extraction. In situations where deterministic, rules-based automation is sufficient, introducing agents can lead to inefficiencies, additional expenses, and possibly even reduced performance.
• As autonomy increases, reliability decreases . As agents gain more decision-making power in workflows, their probabilistic nature can introduce unpredictability, making outputs less reliable and harder to control. It is critical to implement and continually maintain safeguards for agents and to regularly review their authorizations.
• Ethical and practical considerations . Not all decisions should be delegated to AI systems. When agents are used in high-risk or sensitive areas, careful oversight is needed to ensure responsible deployment and prevent unintended consequences.

Given these limitations, it is recommended that before using a proxy, you take the time to reflect on whether it is truly necessary in your specific workflow. The following questions can help you make this judgment:

• Is the task complex enough to require adaptive decision making, or is a deterministic approach sufficient?
• Could simple AI-assisted tools like RAG without agents achieve the same effect?
• Does the workflow involve uncertainty, changing conditions, or multi-step reasoning, and can agents handle these more efficiently?
• What are the risks of giving agents autonomy, and can these risks be effectively mitigated?