A Brief History of Agentic AI: From Transformers to Agent-Native Tooling

Eight and a half years. That is how long it took to go from a research paper published by eight Google engineers to the point where AI agents can autonomously run your business's marketing pipeline, handle your customer inbox, and push code to production without a human in the loop. In the grand sweep of technological history, that is extraordinarily fast. And most businesses have not had a chance to catch up.

This is the story of how agentic AI happened. Not the hype version. Not the breathless Silicon Valley recap. The real story — milestone by milestone — of what changed, why it mattered, and what it actually means for companies in Manchester, Leeds, Birmingham, and everywhere else in the UK trying to figure out whether AI is relevant to them right now.

Spoiler: it is. And understanding the arc of how we got here is the clearest way to understand where things are going next.

June 2017: The Paper That Changed Everything

The story starts in a Google research lab, with a paper that almost nobody outside machine learning circles noticed at the time. "Attention Is All You Need," published by Vaswani et al. in June 2017, introduced the Transformer architecture — the foundational mechanism behind every large language model that exists today, including GPT-4, Claude, Gemini, and Llama.

Before Transformers, AI language models worked sequentially. They processed words one after another, which made it hard to capture long-range relationships in text and very difficult to parallelise training across hardware. The Transformer changed this with a mechanism called self-attention, which let the model consider all parts of a sequence simultaneously, weighing which parts were relevant to which other parts.

In practical terms, this meant AI models could finally understand context. Not just the word in front of them, but the relationship between words paragraphs apart. The quality of language understanding jumped dramatically, and — crucially — the architecture scaled. The bigger you made the model and the more data you fed it, the better it got, at a rate that continued to surprise even the researchers building it.

For businesses, 2017 meant nothing directly. The Transformer was a piece of academic infrastructure, not a product. But it was the foundation on which everything else would be built. No Transformers, no GPT. No GPT, no ChatGPT. No ChatGPT, no agentic AI. The entire arc traces back to this moment.

November 2022: The World Meets an AI

Five and a half years passed between the Transformers paper and the next seismic event on this timeline. That is not because nothing happened in between — OpenAI launched GPT-2 in 2019, GPT-3 in 2020, and Codex in 2021. But these were tools for researchers and developers, not mainstream products. The broader world was not paying attention.

That changed on 30 November 2022, when OpenAI launched ChatGPT.

Within five days it had a million users. Within two months, a hundred million. No consumer application in history had grown that fast. People who had never thought about AI suddenly had an opinion on it. Every newspaper ran pieces about it. Every company started asking their IT teams what it meant for them.

ChatGPT was a chat interface wrapped around GPT-3.5, optimised via a technique called reinforcement learning from human feedback (RLHF) to be helpful, conversational, and to refuse harmful requests. It was not technically the most advanced AI available — GPT-4 came several months later — but it was the first AI product that felt genuinely usable by ordinary people.

For businesses, ChatGPT mostly looked like a very clever writing assistant. You could ask it to draft emails, summarise documents, explain concepts, or generate marketing copy. Useful, certainly. But it was still a tool you typed questions into and read answers from. It could not do things. It could not take actions. It was a chatbot — a phenomenally capable one, but a chatbot nonetheless.

That distinction matters enormously. An AI that answers questions is a productivity tool. An AI that takes actions is an agent. And the step between those two things is not a small one.

June 2023: The Moment AI Learned to Act

The single most consequential technical development in the history of agentic AI is not widely celebrated. It did not generate the headlines that ChatGPT did. Most business owners have never heard of it. But in June 2023, OpenAI added native function calling to their API, and everything changed.

Here is what function calling means. Before it, you could ask an AI a question and it would give you a text answer. After it, you could describe a set of tools to the AI — "here is a function that searches the web, here is one that reads a file, here is one that sends an email" — and the AI could decide to call those functions itself, receive the results, and continue reasoning. For the first time, an LLM could reliably take actions in the world, not just generate words about it.

This was the birth of the practical AI agent. Not in theory. In production. Developers immediately began building systems where AI could browse the web, query databases, call APIs, write and execute code, and send messages — all orchestrated by the model itself based on a goal you gave it.

The shift in what was now possible was profound. A business could, in principle, give an AI model a task — "find me the ten most relevant leads from this database and draft a personalised outreach email for each one" — and the model could execute it end-to-end. No human in the loop. No manual steps. Just a goal and a set of tools.

In practice, it was messier than that. The frameworks for building agentic workflows were nascent. LangChain, which emerged in late 2022 and matured rapidly through 2023, gave developers a way to chain together LLM calls and tool use, but it was complex, brittle, and required significant engineering effort. AutoGPT went viral in early 2023 as a proof of concept for fully autonomous agents, but real-world reliability was poor. Devin and Cursor pushed into AI-assisted coding. The ecosystem was chaotic, experimental, and deeply exciting if you were a developer — and largely inaccessible if you were not.

The fundamental problem was interoperability. Every AI agent framework had its own way of defining tools, its own way of managing context, its own architecture. If you wanted to connect an AI to your CRM, your calendar, your database, and your email, you needed a different integration for every system and every framework. It was powerful but fragile. Scaling it was engineering-intensive work.

2023-2024: The Age of Frameworks (and Why Most Businesses Missed the Wave)

Through 2023 and into 2024, the agentic AI ecosystem exploded with activity. LangChain grew into a comprehensive framework for building LLM applications. AutoGen from Microsoft explored multi-agent collaboration, where multiple AI models could work together on a task. CrewAI offered a higher-level abstraction. Dozens of AI coding tools — Cursor, Codeium, GitHub Copilot, Devin — pushed autonomous AI into software development workflows. Retrieval Augmented Generation (RAG) became the standard approach for giving AI models access to private business data.

For developers with time to experiment, this was a golden age. But for most businesses, it remained inaccessible. The tools required deep technical knowledge to set up, significant ongoing engineering to maintain, and a tolerance for unpredictable behaviour. You could not just buy an agentic AI system off the shelf. You built it — painstakingly, from components that did not always fit together neatly.

The companies that benefited most were the ones with strong engineering teams who could dedicate months to building custom AI pipelines. That meant well-funded startups and large tech companies. The SME trying to run a lean operation, the agency, the consultancy, the local service business — they mostly watched from the sidelines, reading articles about AI transformation that felt like they were describing a different world.

This is the missing middle in most AI history narratives. Everyone talks about the ChatGPT moment and then jumps straight to "AI is everywhere now." They skip over the two years in which agentic AI was real and powerful but accessible mainly to those with significant technical resources. That gap matters, because it explains why so many businesses feel behind — not because they were slow, but because the tools genuinely were not ready for them yet.

Two things changed that. Both came from Anthropic.

November 2024: The Standardisation Moment — Model Context Protocol

On 25 November 2024, Anthropic released the Model Context Protocol, or MCP. It is an open standard — think of it like USB-C, but for connecting AI models to tools and data sources.

Before MCP, every agentic AI integration was bespoke. You wanted your AI agent to access your Google Calendar? You wrote a custom integration. You wanted it to read from your Notion database? Another custom integration. Connect it to Slack? Another one. Each integration was tightly coupled to the specific framework you were using, meaning swapping out your AI model or your orchestration layer required rewriting everything.

MCP solved this with a simple, elegant idea: define a standard protocol that any AI model can use to discover and call tools, and any service can implement to expose its capabilities. If your calendar app implements an MCP server, any MCP-compatible AI agent can use it, regardless of which model it is running or which framework it is built on. The ecosystem becomes composable.

The uptake was rapid. Within months, MCP servers existed for hundreds of popular tools and services. The major AI frameworks added native MCP support. Suddenly, the integration work that had previously taken weeks of engineering could be done in hours by connecting pre-built MCP servers.

For businesses, MCP was the moment agentic AI started to become genuinely accessible. Not because MCP itself was a consumer product — it was still a technical standard — but because it catalysed a wave of tooling built on top of it that dramatically reduced the engineering barrier. The jigsaw pieces that had existed in isolation started to fit together.

October and December 2025: Agent Skills and the Open Ecosystem

Two more milestones close out this timeline, both arriving in the final months of 2025.

On 16 October 2025, Anthropic released Agent Skills in Beta. Where MCP addressed how AI models connect to tools and data sources, Agent Skills addressed something different: how AI agents acquire and share capabilities. An Agent Skill is a modular, portable package of instructions, tools, and context that gives an AI agent a specific set of abilities — like researching leads, writing blog posts, managing a calendar, or handling customer support queries.

The analogy to software packages is useful here. Before package managers like npm or pip, every developer had to write their own version of common functionality. After package managers, you could pull in battle-tested libraries and focus on the unique parts of your problem. Agent Skills does something similar for AI agents. Instead of building every capability from scratch, you can pull in a skill — a unit of agent capability — and extend it for your specific use case.

On 18 December 2025, this went further still with the launch of the Agent Skills open standard at agentskills.io in Beta. The goal: a community-driven ecosystem where skills can be shared, discovered, and improved across the entire agentic AI community, not locked into any single vendor's platform. Think of it as the npm registry, but for AI agent capabilities.

Together, MCP and Agent Skills represent the standardisation moment that the agentic AI ecosystem needed. MCP handles the integration layer — how agents connect to the world. Agent Skills handles the capability layer — what agents know how to do. With both in place, the complexity of building a capable AI agent system drops dramatically.

We are still in the early innings. Both MCP and Agent Skills are in Beta at the time of writing. The ecosystem is young. But the trajectory is clear, and the direction of travel for businesses that want to adopt agentic AI has never been more navigable.

The Arc and What It Means

Lay the timeline out flat and the arc is striking.

The Transformers paper gave AI the ability to understand language at scale. ChatGPT gave that ability a face and brought it into the mainstream. Function calling gave AI the ability to act, not just speak. LangChain and the agent frameworks gave developers the tools to build the first generation of AI agents. MCP gave those agents a standard way to connect to the world. Agent Skills gave them a standard way to share and acquire capabilities.

Each step unlocked something that was not possible before. And each step compressed the timeline further. From Transformers to ChatGPT took five and a half years. From ChatGPT to function calling took seven months. From ChatGPT to MCP took two years. From MCP to Agent Skills took thirteen months. The pace of development is not slowing down.

For UK businesses, there are a few practical conclusions to draw from this.

The first is that the complexity barrier has fallen substantially. In 2023, building a reliable AI agent for business use required a dedicated engineering team and months of work. In 2026, with MCP-compatible tooling and Agent Skills available, the engineering lift is a fraction of what it was. That does not mean it requires no expertise — it does — but the expertise required is now within reach of a competent AI development partner, not just the largest tech companies.

The second is that standardisation is what makes this durable. In the early years of agentic AI, the fear among businesses was that any investment in AI tooling would be made obsolete by the next framework or the next model. With MCP and Agent Skills establishing open standards, that risk is significantly reduced. Skills and integrations built to these standards will work across models and frameworks, meaning the investment has longer-lasting value.

The third is that the agents being built today are meaningfully different from the chatbots of 2022. An agent that can autonomously research leads, draft and send personalised outreach emails, publish SEO content, manage your calendar, and triage your inbox is not a writing assistant. It is a workforce multiplier. For a small business or agency, that changes what is economically possible.

What This Means If You Are a Manchester Business in 2026

The history above is not academic. It is context for a decision many UK businesses are facing right now: is it time to bring AI agents into how we operate?

The honest answer, as of early 2026, is that the technology has matured enough to be genuinely useful for a wide range of business operations — and the cost of getting started has dropped significantly. The fragile, hard-to-maintain agent systems of 2023 have given way to more robust, standardised tooling built on MCP and Agent Skills. Real-world deployments — AI that handles lead research, content publishing, customer communications, and more — are running in production for businesses far smaller than the tech giants.

The catch, as it always is with new technology, is implementation. The tools are better. They still require expertise to deploy well. And "expertise" in 2026 means something specific: understanding how to design agentic workflows, how to select and configure the right models, how to connect them to your existing systems via MCP, and how to build in the right guardrails so agents behave reliably and safely.

That is exactly the work we do at App Web Dev Ltd. We are a Manchester-based AI agency that has been building on top of these exact technologies — from the early framework days through to MCP and Agent Skills. The outreach pipeline that found this post's topic, wrote the research brief for it, and is now delivering this article to you is itself an example of what agentic AI looks like in practice. Not hypothetical. Running.

If you are curious about what an AI agent system could look like for your business — whether that is automating your lead generation, your content production, your customer communications, or something else entirely — we would like to talk. Not to sell you a platform. To understand your specific situation and figure out whether agentic AI can genuinely help.

Get in touch at appwebdev.co.uk. The eight-year journey from the Transformers paper to agent-native tooling has brought AI to your doorstep. The question now is what you do with it.