With Agentic AI, GenAI Is Evolving From Words To Action

---

          Agentic AI enables AI to “act” rather than just “think” — and paves the way for more advanced and versatile general-purpose AI-based apps in fields like automation, personalized services, and intelligent systems. These systems set their own goals, make complex decisions, and adapt to changing environments, showing more humanlike agency. Enterprise technology leaders and product teams building generative AI apps must understand what agentic AI is and how they can create agentic workflows and design patterns in real-world situations. Read this report to learn about agentic AI and how agentic workflows will transform work.

The recent evolution of generative AI (genAI) involves a progression through three stages, each building upon prior advances. GenAI is moving from AI technology that merely interacts with humans through natural language and visual understanding to complex, autonomous, and adaptive AI systems that can manage complex tasks and interactions. The progression of genAI-based design patterns in each stage illustrates the growing sophistication of this emerging technology; they aren’t mutually exclusive but in fact coexist and are often used in combination. The three stages are:

1. Stage 1: prompt engineering and instruction fine-tuning. The first stage centered on prompting and prompt engineering; the primary focus involved crafting and refining input prompts to guide genAI models in order to produce the desired outputs. The art of prompt engineering involves understanding the model’s behavior and tweaking prompts to optimize the quality and relevance of the generated content. This technique leverages the preexisting capabilities of large language models (LLMs) by using carefully designed prompts to elicit specific responses. Prompting and prompt engineering, combined with fine-tuning techniques, have empowered users to interact with foundation models, enabling applications in content generation, customer service, and more.
2. Stage 2: knowledge integration and retrieval. Building on the foundation of prompting, the second stage in genAI’s evolution involves integrating external knowledge sources to enhance the accuracy and relevance of AI outputs. Techniques like retrieval-augmented generation (RAG) combine retrieval mechanisms with generative models to access and incorporate up-to-date and domain-specific information. In this approach, AI systems retrieve relevant documents or data from knowledge stores in vector databases, which inform and ground the generative process. This integration enhances the accuracy and relevance of AI outputs and overcomes LLMs’ tendency to hallucinate. By using external data sources, techniques like RAG allow AI to give more reliable responses — crucial in fields like research and education where precise information is vital.
3. Stage 3: agentic AI systems. In this stage, AI systems will transition from passive tools to autonomous agents capable of perceiving their environment, making decisions, and taking actions to achieve specific goals. Using agentic workflows, these multiagent systems manage and execute complex tasks autonomously in coordination with other agents while adapting to dynamic conditions. This enables sophisticated task execution, decision-making, and collaboration. For instance, genAI-based applications will autonomously handle complex tasks like planning intricate travel itineraries or resolving customer service requests without human intervention or supervision. Agentic AI is a significant leap forward in genAI capabilities, paving the way for more advanced and versatile general-purpose AI-based apps in fields like automation, personalized services, and intelligent systems. Agent systems already significantly outperform simpler LLM prompt strategies (see Figure 1).

While the concept of AI agents has existed since the era of rules-based systems, early agents were inflexible and of limited use. GenAI has enabled more specific design patterns and capabilities for AI agents; they can account for the dynamic, unpredictable nature of real-world interactions. Agentic AI systems are a way to architect AI agents in practice. We define agentic AI as:

Advanced AI systems, powered by foundation models, that demonstrate a high degree of autonomy, intentionality, and adaptive behavior, extending beyond the capabilities of traditional and deterministic AI agents. These systems can plan, make complex decisions, and adapt to changing environments, thereby driving toward the highest levels of autonomy in complex process execution.

Agentic AI leverages the capabilities of LLMs by using primitive reasoning loops to accomplish a task from end to end in ways that weren’t possible with simple prompt strategies. Based on computer scientist Andrew Ng’s work, agentic AI is characterized by four design patterns, to which we’ve added two of our own (see Figure 2):

1. Reflection. This is an agent’s ability to critique its own output and use this critique for iterative improvement. The agent adapts to feedback through self-generated natural language prompts, which refine its outputs over time. Memory forms an important part of the architecture, as it allows state persistence. Reflection can use either the same model for generation and critique or different models for each task: Key aspects include self-critique, iterative improvement, feedback integration, and model flexibility. The models enable agentic systems to assess themselves, continuously improve, and achieve higher-quality performance and adaptability without extensive retraining. For example, an agent tasked with writing essays can refine its work based on self-critique, iteratively improving its coherence, grammar, and relevance. Recent frameworks like ReAct and Reflexion have been widely adopted to support these reflection skills in agents.
2. Memory. Memory and state persistence are a key area of distinction between raw LLMs and agentic AI. By design, LLMs are stateless — that is, they don’t retain details of previous interactions such as state changes. The “memory” in an LLM interaction is limited by the size of the context window and suffers from context pollution. Moreover, a raw LLM has limited ability to glean deeper, longer-term insights from raw data. Agentic systems avoid this problem by adding active memory, which includes short-term memory like inputs from previous tool calls, state information, and other forms of active knowledge and long-term memory including episodic, semantic, and procedural memory. Today’s memory modules typically combine different distributed sources, such as in-memory databases for caching, relational databases for operational data, and vector databases for embedding.
3. Planning. This is an agent’s ability to decompose complex tasks into manageable subtasks and plan and coordinate their execution across models in order to achieve a specific goal. This involves 1) task planning, where AI analyzes the user’s request to understand their intent and decompose it into smaller, solvable tasks; 2) model selection, where AI chooses the most appropriate models for each task; 3) task execution, where AI invokes the selected models to perform subtasks and manages resource dependencies to ensure that the outputs of prerequisite tasks are used correctly; and 4) response generation, where AI integrates the results from all executed tasks to provide a coherent and comprehensive response to the user. In practice, an agentic workflow could have a planner agent that analyzes an incoming prompt, breaks it into components, and coordinates the execution of each component among multiple specialized agents (see Figure 3).
4. Tool use. This is the agent’s ability to employ external software, services, modules, or data sources to extend its functionality beyond its built-in capabilities. It shifts the focus of the AI system from a passive role of content generation to an active role where it executes actions to reach a defined outcome. For instance, a travel-planning AI agent might use APIs to query flight information, autonomously plan an itinerary, and even book flights based on the user’s known preferences. Agents today largely leverage tools like APIs, microservices, knowledge bases, visual encoding models, and other language models. But tool use is a flexible concept and can extend into process or task automation workflows as well as interactions with humans in the loop. By using tools, agentic AI can deliver high-quality outcomes, adapt effectively to diverse real-world applications, and become a key part of an organization’s automation fabric.
5. Multiagent collaboration. Multiple agents allocate tasks, share information, and debate ideas to achieve better solutions than a single agent could. This involves 1) task allocation, where a planning or administrative agent assigns tasks to other agents based on their expertise; 2) communication and coordination; 3) role specialization, with agents focusing on areas like analysis, programming, and testing; and 4) dynamic adaptation, which allows agents to adjust their strategies based on feedback and changes in the environment. In a content creation studio, scriptwriting, editing, and research agents collaborate to produce high-quality content. These specialized agents’ behavior can be defined using natural language system prompts or configuration documents and augmented with external learning and memory retrieval; agents may also engage in a dialogue with each other. This design pattern differs from the planning design pattern, where a single agent breaks down tasks and coordinates their execution in a structured, sequential manner.
6. Autonomy. This is an agent’s ability to initiate and execute actions independently without needing to check in with a human during workflow execution unless specifically required to. Once given a goal, the agent is in the driver’s seat. An early example of autonomy is Devin, Cognition AI’s AI software engineer agent that can learn new technologies, handle external data, test ideas, build apps from start to finish, identify and debug mistakes in its own code, and even train its own AI models. While autonomy is a fundamental characteristic of agentic AI, it is actually a superset of planning, reflection, tool use, and multiagent collaboration: Coupled with active learning and the other five design patterns, it enables sophisticated, autonomous systems to be built from relatively simple agentic “LEGO blocks.” These systems can help solve complex scientific and engineering problems, such as protein modeling, molecular mechanics, autonomous mobility, or mechanical engineering problems.

The idea of agents and agentic workflows isn’t new: For a decade, the public imagination has been captured by deep-learning-based agents from consumer voice assistants like Alexa and Siri to virtuoso AI systems like Deep Blue, AlphaGo, and AlphaFold. Today’s emerging agentic AI differs from these narrower systems thanks to two key properties: They can effectively perform a wide range of diverse tasks and they are programmed and managed via natural language prompts rather than intricate coding. Applying these two properties to the design patterns described here results in three different classes of agentic interactions that will exist in future agentic workflows:

1. Agent-to-agent interactions. Agents will interact in sophisticated ways, collaborating to tackle complex tasks efficiently; each specialized agent will contribute its expertise to achieve common goals. Autonomous agents will dynamically form teams, negotiate roles, and distribute tasks according to their capabilities and the task’s demands; they will iteratively improve their collective performance through continuous planning and reflection. This will enable agentic AI systems to solve multifaceted problems that single agents can’t handle, such as coordinating supply chain logistics or managing disaster response efforts. Advanced protocols will ensure effective data exchange and communication across different platforms, fostering a network of intelligent systems that operate with minimal human oversight. As these interactions evolve, agents will form an interconnected web of intelligent systems that work together to enhance productivity, resilience, and adaptability.
2. Agent-to-human interactions. Agents will significantly augment human capabilities, handling routine and repetitive tasks and enabling humans to focus on strategic, creative, or higher-order tasks. Agents will help doctors analyze medical data to recommend personalized treatment plans and assist financial analysts in identifying market trends and optimizing investment strategies. As agents learn from and adapt to human feedback, they will become more intuitive, fostering a symbiotic relationship. This collaboration will make the workforce more efficient and productive; together, humans and agentic AI will achieve outcomes that neither could accomplish alone. At the same time, consumer-facing AI agents will affect consumer interactions as they make their way into consumer products and devices, shaping the evolution of agent-to-human interactions across the board in ways that are less well understood today.
3. Agent-to-environment interactions. Agentic AI will play a pivotal role in transforming agent-to-environment interactions within a firm’s business ecosystem. Agents will autonomously interact with internal and external business environments, executing processes, making decisions, and optimizing key business functions. Internally, agentic AI will streamline operations by managing supply chains, automating workflow processes, and maintaining IT infrastructure, thus ensuring efficiency and minimizing human error. Externally, it will enhance customer interactions by personalizing experiences, managing supplier relationships, and monitoring market trends to inform strategic decisions. Through advanced planning, reflection, and multiagent collaboration, agents will adapt to changing conditions and seamlessly integrate into business processes. Agents will use real-time data from events platforms or process mining tools to make informed decisions, predict future trends, and respond proactively to challenges; they will integrate with an enterprise’s automation fabric to help deliver the vision of the autonomous enterprise.

We are still at the dawn of the agentic AI revolution. The current rate of development and investment in genAI means that agentic AI will rapidly evolve and find early use cases in various business areas. Pharmaceutical company Novo Nordisk is using AutoGen to develop a production-ready multiagent framework for clinical data analytics. However, much of the technology and many design patterns are still under development and far from reliable, consistent, or valuable. Tech leaders should remember that:

1. Agentic AI is still in the early stages of development. Its full potential and limitations aren’t yet fully clear. Many of the core design concepts are only just making the leap from academic papers to real-world implementations. This raises several challenges: As frameworks mature, we may see problems with standardization and reliability; the rapid pace of AI innovation also means that today’s cutting-edge solutions may quickly become obsolete or be supplanted by better approaches. Moreover, this evolution is happening at a time when foundation models are becoming more powerful and complex, thanks to growing context window sizes and multimodal capabilities. Organizations must temper their interest in AI agents with caution, allowing for iterative improvements and adjustments as the technology evolves.
2. Vendor hype has led to confusion about agentic AI’s true capabilities and limitations. In the strictest sense, agentic AI refers to LLM-based agents interacting in multistep workflows that adhere to the design patterns we describe. However, traditional automation firms are already coopting the term and rebranding their robotic process automation, digital process automation, and conversational automation products as “agentic.” The media is exacerbating this by applying the term “agentic” too broadly. Misunderstanding about what AI agents can achieve with today’s technology, along with poor platform choices, may lead to regrets. Tech leaders should ask someone on their team to keep tabs on the latest academic research on AI agent frameworks; pilot solutions on a small scale before implementing them widely and be prepared to fail fast; and counter excessive genAI enthusiasm and expectations among their nontechnical peers via clear communication about the capabilities and limitations of AI agents.
3. Agent development tools and frameworks are also nascent. A broad range of tools and platforms are being developed to offer various components of agentic AI — from prebuilt special-purpose agents to “build your own agent” platforms. OpenAI’s Assistants API allows users to create agents natively within the GPT ecosystem. LangChain Agents and LlamaIndex Agents allow users to build agents supported by tools, LLM “chaining,” and memory. Other agent development frameworks include AutoGen (an open-source project by Microsoft), AutoGPT, BabyAGI, BCG AgentKit, and Crew AI. Hyperscalers also have early offerings, such as Alibaba Cloud’s ModelScope-Agent; AWS’s Agents for Amazon Bedrock; Baidu AI Cloud’s AgentBuilder; Huawei Cloud’s AppStage; Tencent Cloud’s Yuanqi; and Vertex AI Agent Builder from Google Cloud.

If you’re looking to build or leverage agentic AI at your company, you must ensure that you:

1. Integrate agentic AI into your broader automation strategy and fabric. Relying solely on agentic AI without considering the overall automation landscape can lead to fragmented processes, broken workflows, and unexpected customer outcomes. Effective integration requires careful planning to ensure agents work seamlessly with existing technologies and enhance, rather than disrupt, business operations. Map out how agents will interact with other systems within existing processes and data (see Figure 4). Consider agentic AI as part of a larger strategy that includes other automation tools and processes to ensure a cohesive approach to digital transformation. Successful integration can lead to improved efficiency, reduced costs, and enhanced decision-making capabilities across the organization.
2. Apply strong data guardrails to agentic AI. Agents will be in your processes, interacting with your data. They may handle sensitive information, including customer data, financial records, and proprietary business insights. Protecting this data from unauthorized access and making careful choices when exposing training data to AI is key to maintaining trust and compliance. As with all AI, you must apply a robust set of guardrails and security protocols, including encryption, access controls, and regular audits to detect and address vulnerabilities, identify prompt injection-based attacks, and avoid cascade failures. You must also establish clear data governance policies that dictate how AI agents collect, store, and use data.
3. Place trust at the center of your agentic workflows. Like all forms of AI, agentic AI can inadvertently perpetuate or amplify biases present in the training data, producing unfair and unethical outcomes. GenAI in particular struggles with gaps in explainability, due to a lack of transparency around its internal architecture and reasoning. Implement stringent checks and balances, including bias detection tools, diverse datasets, and inclusive development practices. Understand the biases and issues inherent in the foundation models you choose for various agents. Regular audits of and transparency in AI processes are also essential to identify and mitigate biases. Ethical considerations should guide your agent deployment to ensure they are used in ways that align with your organization’s values as well as societal norms. By proactively addressing bias and ethical concerns, you can prevent harmful outcomes, build public trust, and ensure that your AI agents contribute positively to your business and the wider community.

For the compete Forrester Article Click

https://reprint.forrester.com/reports/with-agentic-ai-generative-ai-is-evolving-from-words-to-actions-9c6cf2d9/index.html

We would like to thank the individuals from the following companies who generously gave their time during the research for this report.

Amazon Web Services

Automation Anywhere

Infosys

Microsoft

NTT DATA

Uber Technologies

UiPath

United Overseas Bank

About Forrester Reprints