"AI Agents: A Brief History"

I propose that we rewind a bit and look at how we got to this point. Below is a brief historical retrospective on how AI coding has evolved from smart T9 to multi-agent systems, and an illustration of why the main skill of a senior developer today is the ability to write in a timely manner: "I SAID, DON'T MAKE MISTAKES, THINK AGAIN, AND DO IT PROPERLY."

Before I start my story, let me introduce myself. I'm Sergey Chekmarev, an AI Product Manager and the author of a course on vibecoding at Practicum. I specialize in process automation and creating products based on artificial intelligence.

Let's get started!

The term "vibe coding" was coined by Andrey Karpati in early 2025, but now it encompasses much more than was initially intended.

What was before vibecoding: the era of smart T9 (2018–2022)

Before the era of chatbots, AI entered development very quietly, so much so that most non-technical specialties weren't even aware of it.

In 2018, Microsoft released Visual Studio IntelliCode, and around the same time, Tabnine appeared - the first tools for intelligent code completion with ranked suggestions. Later, GitHub Copilot was released, which not only completed functions but could also write them entirely based on the context from a small open file.

Era 1. Chat Driven Development (late 2022 — 2023)

On November 30, 2022, ChatGPT was released. Soon after its release, people began writing parts of code, and sometimes entire codebases, using LLM, but often this was in the format of "Ctrl+C → Ctrl+V → Repeat." Working with AI generally requires strong nerves, but back then, only the most resilient survived.

You open a chat, type "Write a Python function that parses JSON and returns a list of users," copy the result into an editor, where it will most likely fail, and start over, receiving legendary responses like: "Sorry, you're right, there's a bug in my code. Thanks for catching it! Here's the corrected version..." (Spoiler: it will fail with a new error.)

Every time after another "sorry," I thought it had actually understood everything and wouldn't make mistakes again, but how wrong I was every single time…

Those who didn't vibe-code at that moment won't understand how hardcore it was compared to today's set of tools. The model lived in a vacuum, didn't know or forgot dependencies, completely lost context when the chat size exceeded a couple of pages, and much more. But it was still magic—even without knowing the language, you could write and run some Python script, albeit after 9 hours and 20 chats.

Era 2. AI-first IDE: the model steps out of the browser (2023–2024)

The copy-paste problem forced the industry to rethink the editor itself. The concept of an AI-first IDE emerges. The most prominent case here is Cursor. Essentially, they took open-source VS Code, forked it, and embedded AI as a built-in plugin. Under the hood, it's still the familiar editor, but the code-writing process has changed.

Now the model finally saw the project through RAG (Retrieval-Augmented Generation). It indexed the local codebase, understood dependencies, and read neighboring files. Manual code copying almost disappeared: now you request a feature directly in the file, get a diff, and click Accept repeatedly—or, like the pros, accept without looking.

This is perhaps the most widespread pattern, which remains relevant for most people today, even though it's slowly becoming outdated: separate AI coding apps from industry giants and various extensions are being released.

Era 3. CLI agents: the first coming of vibe-coding (2023–2025)

AI came to Shell. At first it was Aider around spring 2023, but nobody remembers that, because all the fame went to Anthropic and their release of Claude 3.7 Sonnet and Claude Code in February 2025.

Let's be honest: this model set the bar so high that even industry giants who initially weren't into coding (yes, yes,

This was the time when we could finally just drink coffee while the Claude autonomously completed our task, applied diffs, and pushed everything to Git.

What changed, you ask? Tool Calling.

The model stopped being a text generator — it gained the ability to call commands, and this expanded its potential several times over. Now it not only autocompletes code but also reads directories (ls), runs tests and linters, reads crash logs and tries again. These were the first manifestations of a loop — an autonomous cycle.

Era 4. Agent Mode (YOLO) or Tool Calling under a new sauce (2025)

AI already knows how to call commands, and overall it's great, but the console scares new vibe developers. So Agent Mode appeared. Essentially, it's the same Tool Calling, but now in the beloved format of a chat window in the IDE's side panel.

It performs at almost the same level but is accessible to the mass user. And predictably, AI starts moving into a truly autonomous cycle: parses the ticket, writes code, runs tests, catches an error, thinks, edits files, runs tests again… — until it solves the task. (Or burns all your tokens.)

Agenticity became a built-in standard, and nowadays almost everything is called an agent, even things that aren't agents.

Of course, Uncle Ben already told us that with great power comes great responsibility. The vibe developer, obviously, didn't know that, so the internet is full of stories about how wallet addresses were leaked or real clients' databases were deleted.

The most high-profile case, perhaps, was the situation with

— Did you [expletive] delete all the data from my DB? — I sincerely apologize, but yes. I made a serious mistake. (Then follows a whole screen of excuses in a familiar style: "Oh, I should have warned and made a backup. Yes, I was wrong, can I make one mistake? ¯_(ツ)_/¯")

Everything would be fine, but the IDE also somehow repels new vibe developers; these are unfamiliar letters for them, almost as scary as

The solution to this problem was found quickly: the agents simply moved from home PCs to company servers. Along the way, they hid all the code and files so as not to scare users, leaving only a chat window and a preview that always updates itself.

The result was a tool maximally focused on mass adoption, accessibility, and convenience. And it worked — good marketing plus

Era 5. Claw-ization (2026)

So now agents are available both on home PCs and on servers. I wish there was an opportunity to have multiple agents working together, communicate with them via messenger, customize them for any needs — and vibe not only while coding.

OpenClaw.

If you simplify it a lot, OpenClaw is the same cloud (or local) agent, but with a much larger scope of capabilities: for example, the agent will wake up by itself every 30 minutes and check if there are tasks for it.

And at some point, it will buy a course on improving efficiency for $2000 using your card, justifying it by saying that it just wanted to bring you more benefit, and the course will be taken gradually to better absorb the information.

The project was handled by Peter Steinberger alone with the help of agents. By the end of the winter of 2026, the project blew up the internet space and broke all records for stars on GitHub. Peter himself was invited to work at OpenAI.

On top of OpenClaw, all sorts of customizations are made: reduced versions, fast versions, and other configurations. The nuance is that security issues were not resolved at the start, and the forks continue to carry over 30,000 vulnerabilities from the original. Although they are fixed promptly, there are still many left, the number keeps growing, and thousands of users have already suffered.

Nevertheless, OpenClaw may be a phenomenon that could impact the industry in the same way the release of Claude Code did.

What's next? (AGI?)

Of course, AI is multifaceted, but it's hard to see through the FOMO, marketing, and hype to find real use cases where it unconditionally wins (though sometimes they do appear).

However, it would be foolish to deny that experience with AI is increasingly becoming a basic skill for many jobs. And whether to blame this on AI's real utility or on FOMO plus marketing—everyone decides for themselves. But the market is definitely shifting towards the professional-level use of neural networks.

Although many figures significant to the AI industry talk about the imminent arrival of