Agents - Why Should You Care in 2025?

If you've been paying even a modicum of attention to the AI hype cycle recently, you've no doubt encountered countless articles, videos, and posts breathlessly proclaiming that AI agents are set to hit the mainstream in 2025.

This year is widely considered pivotal due to several converging factors: advancements in large language models, increasing investment in AI infrastructure, and the development of more sophisticated agentic frameworks capable of handling complex, multi-step tasks. These technological strides, coupled with a growing demand for automation and efficiency, are creating an ideal environment for agents to redefine industries and user experiences. From promises of revolutionising industries to enabling entirely new business models, it's hard to ignore the buzz but is it just tech bros moving from crypto to the new buzz word? Well in the short term this may have a grain of truth but long term, this is more than just hype.

The excitement stems from the tangible potential AI agents have to disrupt traditional systems—whether it's streamlining workflows, creating entirely new revenue streams, or fundamentally altering the way users and businesses interact with technology. These possibilities set the stage for a closer examination of the current state and future potential of AI agents. But beneath the noise lies a genuine paradigm shift—one that could redefine how we interact with technology. In this article, I’ll explore what makes AI agents so significant, their current state, and the disruptive potential they hold for the future.

What Are Agents and Agentic Frameworks?

First, let’s get on the same page: what do we mean by AI agents? Broadly, AI agents refer to artificial intelligence systems that act autonomously to achieve specific goals. They make independent decisions and interact with their environment without constant human intervention. At their core, AI agents have ‘agency,’ which means they can:

• Operate independently and initiate actions once given an objective. For example, an AI agent tasked with managing a social media account could autonomously schedule posts, respond to comments, and analyse engagement metrics.
• Adapt their strategies in real-time based on feedback from their environment.
• Design their own workflows and leverage tools to complete complex tasks.

This autonomy is what sets AI agents apart from traditional AI systems or applications that just use prompt chaining. Agents can optimise their performance by selecting the most appropriate AI model for a given task. For example, they might use smaller, efficient models for simple operations while switching to more powerful models like OpenAI's o1 for tasks requiring higher precision and computational capacity. This adaptability not only ensures efficient resource utilisation but also enhances the reliability of the agent in diverse scenarios. Instead of being programmed to follow a strict set of rules, they have the capacity to dynamically adjust their behaviour to meet objectives.

Agentic Patterns

Andrew Ng, a prominent figure in AI, identifies four key design patterns for AI agents, each with unique capabilities. These patterns not only highlight the diversity of agentic workflows but also form the foundation for many practical implementations discussed later in this article. By understanding these patterns, we can see how they manifest in tools like Perplexity, Cursor, and even industry-specific applications in cybersecurity and healthcare.

1. Reflective Agents These agents can evaluate and critique their own work, identifying problems and iteratively improving their outputs. This self-checking mechanism ensures continuous refinement.
2. Tool-Using Agents These agents integrate with external tools, APIs, and systems to gather information and perform actions beyond their internal capabilities. They’re adept at leveraging external resources to complete tasks.
3. Planning Agents Specialised for complex, multi-step tasks, these agents autonomously develop and execute sequences of actions to achieve specific goals. They can adapt to failures, recalibrate, and reroute as needed.
4. Collaborative Agents Often part of multi-agent systems, these agents work in teams, each fulfilling a specialised role. For example, in software development, one agent might act as a developer, another as a designer, and another as a project manager.

Ng demonstrated that even GPT-3.5, when implemented with agentic workflows, could outperform GPT-4 in certain tasks. For instance, GPT-3.5 agents using reflective patterns excelled at iterative problem-solving, while planning agents were able to autonomously design multi-step workflows to achieve specific objectives. This finding underscores the importance of designing agents with the right capabilities, rather than relying solely on the sophistication of the underlying AI model.

State of the Art in 2024

While the potential of AI agents is vast, what are some practical implementations today?

Frameworks for Building Agents

To enable the development of these sophisticated systems, several frameworks have emerged, catering to different levels of expertise and use cases:

• LangGraph: An advanced framework designed for professionals, LangGraph offers granular control over workflows and integration with tools like LangChain. It’s particularly suited for developers seeking to build complex, multi-agent systems requiring robust error handling and human-in-the-loop interactions.
• CrewAI: Aimed at those looking for low-code solutions, CrewAI simplifies the creation of agentic systems without requiring extensive technical expertise. Its streamlined interface makes it ideal for deploying functional agents quickly while maintaining flexibility for customisation.
• AutoGen: Part of the Microsoft ecosystem, AutoGen enables seamless integration with Microsoft tools and APIs. It’s particularly useful for businesses already leveraging Microsoft’s suite of products, allowing for the rapid deployment of agentic systems tailored to enterprise needs.

These frameworks represent the building blocks for AI agents, making it easier for developers and organisations to design systems that align with their unique requirements.

While the potential of AI agents is vast, what are some practical implementations today?

Perplexity:

Perplexity stands out as one of the most successful implementations of an AI agent. Despite its simplicity, it functions as a straightforward three-step agent: it identifies user queries, fetches relevant information, and presents concise, digestible answers. Its success highlights how even simple agentic designs can deliver immense value using excellent execution and mixture of models. However, the landscape is rapidly evolving, and agents are set to become much more sophisticated in their capabilities.

Cursor:

Cursor, on the other hand, represents a far more advanced implementation of agentic AI. While Perplexity demonstrates the potential of straightforward agentic designs, Cursor showcases the evolution toward more complex, domain-specific capabilities. This progression highlights how agents are not just improving but diversifying in their functionality, a key theme in the broader development of AI systems. It has gained significant popularity among coders, offering features that go beyond mere code suggestions. Cursor can help debug, refactor, and even propose alternative solutions autonomously, acting as a collaborative coding assistant. Its sophisticated design showcases the potential for agents to handle complex, domain-specific tasks with high efficiency. These companies are integrating agentic frameworks to enhance user experiences. For instance, Replit’s Agent provides an advanced AI assistant that can perform a range of tasks beyond simple code suggestions. It integrates deeply into the Replit environment, autonomously debugging, writing, and refactoring code, while also assisting with project management tasks. This makes it an invaluable tool for developers, as it leverages agentic capabilities to streamline coding workflows and enhance productivity within a single platform.

Cybersecurity: AI agents are revolutionising this industry by autonomously monitoring systems, detecting threats, and even taking preemptive action against potential breaches. For example, Microsoft’s Security Copilot integrates GPT models with threat intelligence to analyse and respond to cyber incidents, allowing businesses to address vulnerabilities faster and more effectively. Similarly, tools like Darktrace use AI agents to identify unusual patterns in network activity, flagging potential threats before they escalate.

Healthcare: In the healthcare sector, agents like Google's Med-PaLM 2 are helping professionals interpret complex medical data, such as imaging results or patient records, to improve diagnostics and treatment plans. Autonomous agents are also being deployed for administrative tasks, such as scheduling and billing, which frees up valuable time for patient care.

Legal Tech: In legal technology, AI agents like Casetext’s CoCounsel assist lawyers by sifting through vast amounts of legal documents to identify precedents or draft contracts. By automating tedious tasks, these agents allow legal professionals to focus on strategic decision-making and client interactions.

What Are the Possible Futures for This Technology?

1. Agents as the New Apps

In the same way mobile apps transformed the way we interact with the internet, agents could redefine software experiences. Instead of clicking buttons or navigating menus, users could simply state their goals, and agents would handle the rest—whether that’s booking a trip, conducting research, or managing workflows.

2. Results as a Service (RaaS)

Dharmesh Shah’s concept of Results as a Service (RaaS) builds on the idea that customers increasingly value outcomes over tools. The core processing steps of these systems reveal how they deliver value:

• Query Understanding: Using advanced Natural Language Processing (NLP) algorithms, agents interpret user queries by analysing context and intent. They employ techniques like chain-of-thought prompting and few-shot learning to guide behaviour effectively.
• Information Retrieval: Agents pull real-time data from a variety of web sources. By connecting to external models and leveraging APIs, they retrieve relevant information efficiently. Summarising search results with inline citations ensures the information is both concise and reliable.
• Response Generation: After gathering information, agents create comprehensive, personalised summaries. By applying fine-tuning methods optimized for both accuracy and speed, they deliver context-aware results tailored to user needs.

This transition from SaaS to RaaS is made possible by AI agents capable of executing these multi-step tasks autonomously. For example, instead of relying on accounting software, a business could instruct an agent to complete tax filings or generate financial reports seamlessly. The ability to focus on delivering actionable outcomes instead of tools reflects a paradigm shift in software delivery models.

Results as a Service (RaaS) builds on the idea that customers increasingly value outcomes over tools. Most customers want to buy a hole, not a drill. Rather than providing a platform or software, companies could deploy agents that deliver direct results. For example, instead of using accounting software, a business could rely on an agent to prepare tax filings autonomously. This evolution from SaaS to RaaS is made possible by AI agents capable of executing multi-step tasks independently.

3. The Agentic Web

The web itself could transform into an ecosystem designed for agentic interaction. Imagine a future where people primarily use apps like Perplexity to search for personalised answers, bypassing traditional web pages entirely. Perplexity’s ability to process user queries, retrieve relevant information, and generate concise, context-aware responses demonstrates how agents can offer a seamless user experience. If this trend continues, users may no longer visit web pages directly, instead relying on agents to deliver distilled and actionable insights.

What happens when content creators adapt to this shift? Instead of optimising content for search engine crawlers, major content producers might start deploying their own agents to serve information directly to other agents. This would mark a radical departure from the current model, where agents scrape web pages to retrieve data. By creating agent-to-agent communication networks, content producers could retain greater control over their data and monetisation models.

This shift could disrupt industries like SEO and affiliate marketing. Traditional methods of optimising content to manipulate rankings or drive traffic would lose relevance in an agent-driven ecosystem. Affiliate sites, which often rely on traffic from search engines, might struggle as agents bypass them in favour of direct interactions with authoritative sources.

For Google, the implications are profound. The company’s search business, which heavily relies on ad revenue from search engine traffic, would face significant challenges. As users turn to agents for personalised, actionable answers, the need for traditional search engines could diminish. Google would need to pivot toward creating or integrating with agentic systems to remain competitive. This could involve developing proprietary agents or transitioning its advertising model to fit within the agentic web ecosystem.

The agentic web represents both a technological leap and a strategic inflection point for industries dependent on traditional web paradigms. The question remains: how quickly will businesses adapt, and who will lead the charge in shaping this new digital landscape?

The web itself could evolve into an ecosystem designed for agentic interaction, with projects like Mariner offering a glimpse of this future. As a Chrome extension, Mariner enables agents to autonomously navigate and interact with websites. For instance, it can process forms, click through workflows, and even manage intricate tasks like booking services or compiling research—all without human input. Rather than merely retrieving data, agents like Mariner represent the shift toward enabling full task automation, hinting at how deeply integrated these systems could become in everyday online activities.

What Will This Look Like in Enterprise and B2B?

For enterprises, agents could become indispensable tools, particularly in areas requiring extensive research and analysis. Imagine jobs like market research, policy analysis, or even scientific discovery being transformed by agents capable of autonomously gathering data, synthesising insights, and presenting actionable recommendations. These agents could act as knowledge workers, reducing the time and effort required for high-impact decision-making.

This shift opens up intriguing possibilities, such as the creation of marketplaces or "LinkedIn-style" platforms for worker agents. These platforms could serve as hubs where agents showcase their capabilities, such as specialised research, customer support, or logistics optimisation, allowing businesses to "hire" agents tailored to their needs. Beyond that, agents themselves might use these marketplaces to shop for other service agents, fostering a network of collaboration and specialisation.

The agent-centric approach could redefine traditional roles, streamline operations, and pave the way for entirely new business models. In many ways, this mirrors the promise of Results as a Service (RaaS), where the focus shifts from delivering tools to delivering outcomes. By leveraging agents capable of autonomously managing complex tasks, businesses can align more closely with the RaaS paradigm, creating opportunities for innovation and customer-centric value. The question for enterprises is not just whether to adopt agents but how to strategically position themselves in this rapidly evolving landscape.

For enterprises, agents could become indispensable tools. In B2B settings, they might act as virtual employees, handling routine tasks like generating reports, processing invoices, or managing client communications. Some potential applications include:

• Customer Support: Agents can act as advanced chatbots that resolve complex queries without escalating to human representatives.
• Supply Chain Management: Planning agents can optimise logistics, monitor inventory levels, and reroute shipments in real time.
• Compliance: Reflective agents could continuously audit business practices, ensuring adherence to evolving regulations.

A New Era of Autonomy

The rise of AI agents represents a shift in how we think about automation, software, and business processes. However, it’s important to temper enthusiasm with caution. While AI agents promise to enhance workflows and efficiency, overhyping their capabilities could lead to unrealistic expectations. Challenges such as ethical considerations, privacy concerns, and the technical barriers to seamless integration must not be overlooked.

To prepare for this change, businesses should start experimenting with agent-driven workflows in small, focused areas. These steps align closely with the earlier predictions about agentic disruption, where the ability to manage complex tasks autonomously plays a pivotal role. By integrating agents into their operations, companies can position themselves at the forefront of this transformation, embracing the innovation and efficiency promised by agentic systems. Identifying repetitive tasks or workflows that could benefit from autonomy and adaptability is a key first step. Partnering with AI providers or integrating existing agentic tools can also give businesses a competitive edge in adapting to this new era of digital transformation.

Agents are evolving into collaborators rather than mere assistants. For instance, future workplaces might see agents acting as teammates, augmenting human potential while personalising interactions to make customers feel genuinely understood and valued. The shift from "there’s an app for that" to "there’s an AI for that" signals a profound change in how organisations will leverage these tools. Marketplaces are emerging, enabling businesses to "hire" specialised agents for tasks such as writing meta descriptions, conducting research, and more. At the same time, low-code platforms are democratising agent creation, empowering non-programmers to develop agents tailored to their unique needs.

In these hybrid workplaces, agents will fundamentally change collaboration, acting not just as tools but as integral teammates. Personalisation will no longer mean adding a name to an email but making every customer feel seen, understood, and valued. Despite their intelligence, agents will remain inherently human-centric tools, amplifying what teams can accomplish while maintaining a distinctly human touch.

By combining autonomy with adaptability, agents have the potential to reshape industries and redefine customer expectations. Whether as collaborative partners in the workplace or as the backbone of RaaS models, their role in 2025 and beyond will be pivotal. The real question is: are we ready to embrace this change?