Web 4.0 - The Rise of the Agentic Web
We will need to rethink digital marketing
As artificial intelligence continues its rapid evolution, we're about to see a fundamental shift in how technology facilitates all of our digital experiences. The 'Agentic Web' is a new phase in the evolution of the World Wide Web and will impact how we interact with and through online platforms.
At its core, this shift will be driven by the emergence of digital AI agents: autonomous systems designed not just to assist but to act on our behalf. While the technology industry has long promised various visions of automated assistance, today's convergence of large language models, large reasoning models, and a maturing API ecosystem is starting to make true autonomous digital agents technically feasible.
These developments will have a big impact on the future of digital marketing, challenging our assumptions about how brands connect with consumers. This transition will necessitate not just tactical adjustments but a comprehensive rethinking of digital marketing strategies in the emerging agent-mediated world.
WEB 4.0
The evolution of the World Wide Web has been marked by three distinct eras, each fundamentally transforming how humans interact with digital information and with each other. To understand where we're heading, it's crucial to recognise the patterns of this evolution:
Web 1.0 (1989-2004) represented the internet's first incarnation as a one-way publishing platform. Its static pages and limited interactivity seem quaint by today's standards, but they laid the fundamental infrastructure for global digital connectivity.
Web 2.0 (2004-2020) ushered in the era of social participation and user-generated content. This period saw the rise of platforms that transformed passive consumers into active participants, fundamentally changing how information and influence flow through digital networks.
Web 3.0 (2014-present) introduced the promise of a semantic, decentralised web. While still evolving, this era has been characterised by efforts to make the internet more machine-readable and user-controlled through blockchain technologies and sophisticated data structures.
Now, we're witnessing the emergence of Web 4.0 — which I'm calling the Agentic Web. Web 4.0 will be fundamentally different from previous eras as rather than simply changing how we interact with technology, it will introduce new actors into the digital ecosystem: autonomous digital AI agents that can understand, decide, and act on our behalf.
The Five Levels of AGI
It's important to not only define AI Agents but to also understand where they fit into the broader AI ecosystem that is currently evolving at pace. OpenAI have provided a valuable framework for understanding this progression, outlining five distinct levels of Artificial General Intelligence (AGI) capability that lays out how the technology is likely to develop over the coming years:
Level 1: Conversational AI
Current AI systems excel at natural language interaction but operate within carefully constrained parameters. These systems fundamentally remain reactive tools rather than proactive agents. Their primary value lies in augmenting human capabilities rather than replacing human agency.
Level 2: Reasoning AI
We're now starting to see the emergence of systems that can engage in sophisticated problem-solving comparable to human experts. These large reasoning models are a crucial step toward true AI agents, able to create plans and perform advanced problem solving.
Level 3: Autonomous AI
At this level, AI systems will maintain extended independent operation, managing complex tasks and adapting to unexpected situations without constant human oversight. This capability mirrors the trust we place in human employees to manage ongoing responsibilities.
Level 4: Innovating AI
Level 4 introduces systems capable of not just executing tasks but improving processes and developing novel solutions. This will deliver a fundamental shift from systems that simply follow rules to those that can identify and implement better ways of achieving their objectives.
Level 5: Organisational AI
The final level in this framework defines systems capable of managing entire organisational functions, coordinating complex networks of tasks while maintaining strategic alignment. While this may seem distant, early indicators suggest we're moving more rapidly toward this capability than many anticipated.
Right now we are at an incredibly significant inflection point. As we start to progress from AI models that can reason (level 2) and power more autonomous AI (level 3) we are starting to see the emergence of digital AI agents which will usher in Web 4.0. This transition is actively unfolding as large reasoning models like o1, o3 and R1 begin to power more autonomous systems capable of navigating our complex digital ecosystems. This is where AI will transition from being a tool we deliberately engage with to an active (and eventually pro-active) participant in our digital experiences.
The AI Agent Maturity Framework
What we're starting to see at the beginning of this year is the development and release of specialised digital AI agents, built on top of level 2 large reasoning models. This is also giving us a clear path for how we can move from specialised systems toward broader, more capable, and more autonomous agents. As specialised digital AI agents master specific domains, the technologies developed and learnings made will pave the way for more general-purpose digital agents in the future.
However, the challenge with writing about digital AI agents comes from the fact that collectively we currently don't have a clear definition or agreement of what digital AI agents are. Broadly, I classify digital AI agents as:
"AI systems that are designed to perform digital tasks by interacting with digital ecosystems either under human guidance or autonomously"
But, as OpenAI have done with Artificial General Intelligence, it's useful to break this down into a framework that can help us get a more nuanced understanding of the technology, where our current capabilities are, and where they are likely to take us next. I outline a suggested framework for this below:
The Seven Levels of Digital AI Agent Maturity
Level 1 – Assistive DIGITAL AI Copilots (Specialised, Manual)
Task Scope: Assist with small, well-defined parts of a larger digital task, such as generating code snippets, auto-completing sentences, or suggesting refinements to existing work.
Prompting: Require careful, iterative human prompting to guide responses.
Autonomy: Minimal—functions purely as an advanced suggestion engine.
Examples: GitHub Copilot, Grammarly.
Level 2 – Task-Oriented DIGITAL AI Assistants (Broad, MANUAL)
Task Scope: Handle a broad range of individual digital tasks, such as drafting emails, writing creative content, summarising articles, or performing research queries.
Prompting: Still requires detailed human direction, often in an iterative manner.
Autonomy: Executes tasks based on well-structured user prompts but lacks independent reasoning or planning.
Examples: ChatGPT-4o, Claude 3.5 Sonnet, Google Gemini 2.0 Flash.
Level 3 – Specialised Semi-Autonomous Digital AI Agents (Specialised, Semi-Auto)
Task Scope: Execute highly specialised digital tasks within a specific domain, such as conducting targeted research, analysing financial reports, or autonomously troubleshooting code.
Prompting: Needs an initial prompt but can operate semi-autonomously with limited human intervention.
Autonomy: Can take a single task from start to finish but still relies on predefined methods and lacks adaptive decision-making.
Examples: Devin, specialised Deep Research agents from Google DeepMind and OpenAI.
Level 4 – Generalist Semi-Autonomous Digital AI Agents (Broad, Semi-Auto)
Task Scope: Can complete a broad range of individual digital tasks by interacting with various systems in ways that mimic human input, such as automating research, booking appointments, or handling customer support tickets.
Prompting: Requires a starting prompt but then works semi-autonomously with occasional user input for verification or decision-making.
Autonomy: Has limited reasoning abilities and can navigate pre-defined workflows but struggles with unexpected scenarios.
Examples: In research preview/beta testing: Google DeepMind’s Project Mariner, OpenAI’s Operator, Anthropic's Computer Use, and Apple's Siri enhanced with Apple Intelligence
Level 5 – Multi-Task Digital AI Agents (Multi-Agent Systems, Auto)
Task Scope: Manage multiple digital tasks that are components of a larger process, coordinating different AI systems and APIs like an automated project manager.
Prompting: Needs an initial goal definition but can execute and coordinate multiple subtasks autonomously.
Autonomy: Can dynamically adjust plans and distribute workload across multiple specialised AI agents but still relies on user oversight for major decisions.
Examples: Still theoretical — no real-world implementations yet. In research: OpenAI’s Swarm framework
Level 6 – End-to-End Digital AI Agents (Early Agentic AI, Auto)
Task Scope: Oversee and execute every digital task within a process.
Prompting: Requires only high-level objectives; operates autonomously.
Autonomy: Capable of handling unknown variables and making real-time adjustments without human guidance.
Examples: Still theoretical — no real-world implementations yet.
Level 7 – Fully Autonomous Digital AI Agents (General Agentic AI, Auto)
Task Scope: Manage all tasks of a larger digital process, adapting in real-time to changes in the ecosystem without any human prompting.
Prompting: Require no starting prompt — proactively assesses, prioritises, and executes actions based on context and learned behaviours.
Autonomy: Self-directed, able to set its own goals, make complex decisions, and continuously refine its approach without human intervention.
Examples: Still theoretical — no real-world implementations yet.
This framework should be a useful guide when discussing digital AI agents. By mapping capabilities across seven distinct levels — from assistive digital copilots, to today's specialised and semi-autonomous agents, through to fully autonomous agents — we can better understand what this technology is and how its likely to develop.
As large reasoning models mature and agentic architectures evolve, we're starting to see the first genuine examples of semi-autonomous agents. It's likely that we're entering a period of accelerated development of digital AI agents in 2025, powered by increasingly capable large reasoning models.
Current State of AI Agent Technology
In early 2025 we've already seen the launch of level 3 specialised digital agents as well as early looks at some interesting level 4 general-purpose digital agents. The introduction of Deep Research agents from both Google and OpenAI, alongside Devin's breakthrough in autonomous software development, have shown us real world implementations of level 3 agents that are already incredibly useful and adding a lot of value to many people's workflows. Meanwhile, the research previews and beta releases of Google DeepMind's Projects Mariner and Astra, OpenAI's Operator, and Anthropic's Computer Use are starting to show us what level 4 general-purpose agents could be like.
The progress that we're seeing is driven by advances in large reasoning models, enhanced by multimodal understanding. It is fundamentally changing how artificial intelligence interfaces with our digital ecosystems. Below is a summary of the current state of AI agents:
Level 1: Assistive Digital AI Copilots
When GitHub Copilot was fully released in June 2022 it was the first AI agent to ever be released, operating as an advanced suggestion engine within coding workflows. While sophisticated in its code generation capabilities, it could only generate simple, small snippets of code when prompted by a human.
Level 2: Task-Oriented Digital AI Assistants
The large language models we have seen since ChatGPT was launched in November 2022, including the latest frontier models like ChatGPT-4o, Claude 3.5 Sonnet, and Gemini 2.0 Flash are general models that can perform a broad range of simple tasks, but still require prompting by a human and often have to be used in an iterative workflow.
Level 3: SpecialiSed Semi-Autonomous Digital Agents
We've recently started to see a new generation of specialised agents which have narrow, specialised capabilities but using them is less iterative than working with level 2 agents: Some examples of these newer, level 3 agents are listed below:
- Devin is a coding agent that has some autonomous software development capabilities. It can independently plan and execute a task from start to finish and is able to self-correct and problem solve along the way.
- NotebookLM is a simple, personalised research assistant that is able to analyse documents, YouTube videos, and audio files to produce summaries and find interesting insights. It can also produce an audio overview that turns its output into podcast-like deep dive discussion.
- Deep Research from both Google DeepMind and OpenAI are more sophisticated research planning and execution agents able to semi-autonomously search the web, retrieve relevant content and to synthesise it into a comprehensive research report.
Level 4: Generalist Semi-Autonomous Digital Agents
Over the last six months there have been research previews and beta releases of more general, semi-autonomous agents from Anthropic, OpenAI, Google DeepMind and Apple. This is where the frontier is right now when it comes to AI agents. Below is a brief overview of these level 4 generalist agents:
- Apple's Siri enhanced with Apple Intelligence was previewed at WWDC in June 2024. Apple's vision for Siri is a very integrated approach, embedding agent capabilities directly into Apple's ecosystem. Rather than operating through a browser, it works at the system level to help with tasks across applications while maintaining Apple's focus on privacy and on-device processing. It is expected to launch in Q2/Q3 2025.
- Anthropic's Computer Use was introduced as a public beta in October 2024. It enables Claude to operate computers like a human would - interpreting screen contents, moving a cursor, and typing text. The public beta shows Claude's work in real-time through a dedicated browser window and requires supervision for most actions.
- Google DeepMind's Project Mariner was released as a research prototype in December 2024. It uses Gemini 2.0 to understand and interact with web interfaces, similar to Computer Use but just focssed on web browsing. It moves quite slowly between actions, but achieved 83.5% on the WebVoyager benchmark.
- Google DeepMind's Project Astra was released as a research prototype in December 2024. It is similar to Project Mariner in that it's a general purpose agent, but focuses on real-world assistance through phones and prototype glasses as opposed to just being focused on computer use and web browsing. It can understand voice commands, interpret visual information, and maintain contextual conversations. Project Astra demonstrates how AI agents might integrate more naturally into daily life rather than just automating web tasks.
- OpenAI's Operator was released as a research preview in January 2025. It is an experimental browser automation agent, similar to Computer Use and Project Mariner that can navigate websites and perform tasks like online shopping, travel booking, and research.
Common Traits & Limitations
When looking at the current crop of digital AI agents, and especially the newer level 3 and 4 agents, there are some common traits that they all share.
- The more sophisticated digital agents are built on top of large reasoning models that have more advanced multi-step planning capabilities.
- These models all have large context windows, enabling them to maintain coherent understanding across complex, extended interactions.
- All the level 3 agents are text-based whereas all the level 4 agents are multimodal.
We are yet to see the emergence of true multimodal large reasoning models and I expect that this what we will need before we see true level 4 agents that are ready for prime time.
Despite the significant progress that we're seeing with AI agents, there are still common challenges that exist. The main one is reliability, especially amongst the level 4 agents we're starting to see. I think that this is the main barrier that currently sits between the semi-autonomous agents we're currently seeing and a future state of more autonomous agents. The reality is that until agents are 95%+ reliable at 95%+ of tasks it's very difficult to see how people will trust them enough to perform tasks autonomously.
The other major limitation holding back AI Agents from reaching level 5 and beyond is integrations. I've written extensively about this topic before in my Beyond Chatbots series. The reason why we're seeing level 4 agents that 'operate computers like humans do' is because there isn't a good alternative for AI agents at the moment. The reality is that 'operating computers like humans do' is inefficient, prone to error, and increases the change of AI agents coming up against problems that they can't solve on their own. I see this current trend of 'computer use' as a short-term stop-gap and longer term a fall-back mechanism for AI agents when they come across platforms where they can't interact with it via an API.
To solve this integrations challenge there needs to be a big lift to build out API ecosystems to make them more suitable for integration with AI systems. There needs to be more APIs for more platforms with more read-only APIs that allow AI agents to access data and more write-access APIs that allow AI agents to take actions on behalf of their users. I believe this will come in time, as more people adopt AI technologies it will increase demand for a more mature API ecosystem and more incentives for platform owners to build out their APIs. Only time will tell.
Future Trajectory
So where do AI agents go next? There are some areas that I have a high degree of certainty will improve in the short term, and there are some areas that I am confident will improve, but will take a bit of time.
In the short term (i.e. later this year) we will absolutely see more advanced and more capable text-based large reasoning models that power AI agents. We're highly likely to see large reasoning models with more multimodal capabilities and I'm also expecting to see larger context windows and longer-term memory that will enable digital AI agents to plan better and execute more complicated workflows. We'll also see reduced latency amongst the level 4 agents we've started to see and I expect them to exit research preview/public beta and be more widely released. So I'm confident that we will see fully released level 4 AI agents before the end of the year.
The area that is going to take a bit of time to improve is the API ecosystem. As things stand, I think the Pareto principle (80/20 rule) is at play. There are 20% of digital platforms who could enhance their APIs for digital AI agents which would cover 80% of use cases. The issue we're likely to see unfortunately is that these 20% of digital platforms are the 'walled gardens' (Google, Meta, Apple, Amazon etc.) that are least incentivised to open up their APIs to third party digital AI agents as they are developing their own AI systems.
With the other 80% of digital platforms, it's just going to take time. They will have bigger incentives than the 'walled gardens' to build out their APIs for digital AI agents, but there's just a larger volume of them, and each of them is smaller and will therefore have less of an impact. Interestingly, I think this is where OpenAI and Anthropic could play a big role by working with these digital platforms to help them build out their API infrastructure. This could work in a similar way to how OpenAI is partnering with premium publishers and incentivising them to share their content with their models.
The Impact on Digital Marketing
The emergence and adoption of digital AI agents is going to fundamentally change how brands connect with consumers online. Digital marketing has always been inextricably linked to the development of the web itself, with each new era requiring big changes in how brands promote themselves. Web 4.0 presents what could be the biggest shift yet in how digital marketing operates and will require a comprehensive rethinking of digital marketing strategies.
With Web 1.0, websites essentially functioned as digital brochures, and digital marketing was focused on one-way communication. This meant that brands could easily transfer their traditional broadcast marketing models online. When Web 2.0 emerged in the mid-oughts, social platforms transformed marketing into more of a two-way conversation with user-generated content becoming more central to central brand narratives. We haven't seen Web 3.0 technologies have as much of a direct impact on digital marketing as previous eras but its more decentralised nature has led to the rise of cross-platform identity tracking, data-driven programmatic advertising, and the increasing importance of first-party data and associated privacy considerations.
We're now entering Web 4.0, the Agentic Web, where digital marketing will be defined less by searches, clicks, and visits and more by AI-mediated consumer interactions, the need for API-first marketing architectures, and increasingly human-not-in-the-loop purchase decisions. As digital AI agents become more prominent and are more widely adopted by consumers over the next couple of years we are going to have to get used to the idea of seeing less human traffic online. This will result in fewer human web searches, fewer links being clicked on, and fewer website visits by humans. These have long been the staple of digital marketing strategies and the reason why I believe that Web 4.0 will present the biggest shift yet in how digital marketing operates.
Instead of digital marketing strategies being dominated by human metrics such as searches, clicks, and visits we will see the increasing importance of strategies that cater not just for humans, but for agents and the large reasoning models that power them. When considering large reasoning models, digital marketers will need to ensure that their marketing content is well represented in their training data and develop new techniques for optimising this against their competitors. There will also need to be an evolution of SEO strategies to optimise marketing content for digital AI agents searching the web and direct-to-consumer brands will need to think about how they show up in the decision making processes of digital AI agents. Part of this will be the development of agent-first APIs as we see an increase in human-not-in-the-loop purchases.
Digital marketers will also need to develop new approaches to measurement and analytics as searches, clicks and visits become less important and the 'purchase funnel' as we know it becomes obsolete. Macro modelling techniques like Econometrics should fair well, but more granular approaches like Attribution will need rethinking from the bottom up. Instead of brands optimising digital activity based on granular metrics as they have for the last decade, decision-making will need to become more strategic and sophisticated. Digital marketers will need to understand not just what digital AI agents do but why they do it - similar to how SEO evolved from simple keyword optimisation to understanding search intent and semantic relationships.
The emergence of AI agents means we need to think not just about measurement but about how to structure and present marketing content in ways that agents can reliably process and evaluate. This means considering everything from API design to data architecture to content structure through a new strategic lens.
Preparing for the Agentic Future
The transition to Web 4.0 represents a fundamental shift in how consumers will interact with our digital ecosystems. Brands that want to thrive in the Agentic Web will need to start preparing now, developing both the technical capabilities and strategic mindset required for an agent-mediated digital ecosystem. It is going to require a rethink on how we approach digital engagement at every level.
At a foundational level, digital marketers will need to evolve the technical infrastructure of their online presence. However, it won't just be about creating API endpoints for digital AI agents to access but also about reimagining a brand's entire digital presence through the lens of large reasoning models and digital AI agents. This means developing both new content and new data strategies that can simultaneously support human engagement, the training of large reasoning models, and digital AI agent interactions.
The key to success for digital marketers will be maintaining a balance between optimising for human engagement and developing new capabilities for large reasoning models and digital AI agent interaction. The reality is that human-centric digital marketing will remain crucial even as agent-mediated interactions grow in importance. Brands will need to expand existing digital marketing strategies, not replace them.
Just as the move from Web 1.0 to Web 2.0 didn't eliminate the need for effective websites but added social capabilities on top, the transition to Web 4.0 won't eliminate human-focused digital marketing but will add new AI-focused layers to it. Success will require sophisticated strategies that can simultaneously serve human users, large reasoning models, and digital AI agents effectively.
The key is to start experimenting with these parallel approaches now, learning from early implementations while maintaining excellence in traditional channels. Web 4.0 isn't about replacing the old with the new - it's about broadening our conception of what digital marketing can and should be.
As we look ahead, the Agentic Web will likely bring opportunities we can't yet imagine. The combination of increasingly sophisticated large reasoning models, more capable digital AI agents, and evolving API ecosystems will enable new forms of digital marketing that go beyond our current thinking. The brands that thrive will be those that maintain the agility to capitalise on these opportunities while ensuring their core, human-centric digital marketing capabilities remain strong.
“The future is already here, it’s just not evenly distributed.“
William Gibson