In Issue 7, we watched language models learn to write code. You could type a question into a chat window, and an LLM would hand you a function, a class, an entire module. Extraordinary.

Between 2021 and 2025, that wall came down. AI went from suggesting code to writing, running, testing, debugging, and fixing code — all on its own. The advisor left the room and sat down at the keyboard.

On June 29, 2021, GitHub announced Copilot as a technical preview. Built on OpenAI Codex, a 12-billion-parameter language model fine-tuned on billions of lines of public code, Copilot predicted not just the next word — but entire functions.

Copilot became generally available on June 21, 2022, at $10/month. By February 2023: over one million paying subscribers. An internal study claimed developers completed tasks 55% faster. Critics noted the study used a simple task — but nobody disputed the feeling. Developers called it a “before and after” moment.

“If you read a thousand cookbooks and then write your own recipe, is that copying? What if your recipe happens to match someone else’s word for word? At what point does ‘learning from’ become ‘copying’?”

On November 30, 2022, OpenAI launched ChatGPT. It reached 100 million monthly active users by January 2023 — the fastest-growing consumer application in history. Developers discovered it was shockingly good at writing code from plain English descriptions.

The AI could write code — but it could not run code. It could not see your project. Every interaction followed the same exhausting loop: describe, get code, copy, paste, run, hit error, copy error, go back, get fix, copy again… The human was the bottleneck.

The third leap changed everything: tool use. A language model is a text generator. It cannot read files, run commands, or check whether its code works. It is a brain in a jar — brilliant, but disconnected from the world.

In June 2023, OpenAI introduced function calling for GPT-3.5 and GPT-4. The Toolformer paper (Feb 2023) showed LLMs could learn when to use tools. Anthropic introduced tool use for Claude models in early 2024, generally available with Claude 3 in March 2024. Tool use became a universal pattern.

In July 2023, OpenAI released Code Interpreter for ChatGPT. For the first time, a mainstream AI product could write Python code and run it in a sandboxed environment, see the output, and iterate. The wall between “thinking” and “doing” had cracked open.

Imagine a detective at a crime scene. She thinks: “The window is broken from the inside.” She acts: dusts for fingerprints. She observes: “No prints on the glass.” She thinks again: “So the intruder wore gloves. Let me check the back door.” Think, act, observe, repeat. That is how real problem-solving works.

Before ReAct, there were two separate approaches. Chain-of-thought asked the model to “think step by step” — good reasoning, but no real-world actions. Action-only had the model call tools directly without explaining its reasoning — things got done, but the model often took wrong turns. ReAct combined them.

Aider (Paul Gauthier) pioneered the “AI in the terminal” approach. Cursor (Anysphere, 2023) was an AI-native editor forked from VS Code. Devin (Cognition Labs, March 2024) was announced as “the first AI software engineer” — then faced backlash when demo claims were found overstated. Windsurf (Codeium, late 2024) offered integrated agentic flows. Claude Code (Anthropic, early 2025) operates directly in the developer’s terminal.

A crucial difference: where the AI runs. Sandboxed tools run in locked containers — safe, but the AI cannot access your real project. Terminal-based agents run in the user’s actual environment, with real permissions. More capability, but also more responsibility.

A context window is the maximum amount of text a language model can process in a single interaction. Think of it as the model's working memory — everything it can “see” at once. If information is outside the window, the model cannot access it. It is gone.

That sounds enormous. Surely 200,000 tokens is enough? Not quite. And the reason is a subtle, well-documented phenomenon that researchers call “Lost in the Middle.”

For a coding agent, this creates a slow-motion disaster. The agent starts with a clean, clear context. But as it works — reading files, running commands, analyzing errors — the context fills up. The original plan slides into the foggy middle. The agent may re-do work it already completed, contradict earlier decisions, or “forget” critical constraints.

Here is what separates an agent from a code generator. A code generator writes code. If it fails, you fix it. An agent writes code, runs it, sees it fail, reads the error, reasons about the cause, writes a fix, and runs it again. The AI is active. It fixes its own mistakes.

But error recovery is not unlimited. Agents can fall into repair loops — trying a fix that introduces a new error, fixing that but re-introducing the original, oscillating indefinitely. Good systems detect these loops: after three similar failed fixes, the agent steps back and tries a fundamentally different approach.

A single agent, no matter how capable, is trying to hold an entire project in one context, work on it sequentially, and maintain perfect recall across thousands of steps. That is not how humans build complex things either. Humans build complex things with teams.

“Think about the hardest project you have ever worked on. Could one person have done it alone, no matter how talented? Or did it require different people with different skills, working in parallel, reviewing each other’s work? Now think about why that same logic applies to AI.”

In 2021, Copilot gave developers a taste of AI-assisted coding — magical, but fundamentally autocomplete. In 2022, ChatGPT proved you could describe what you wanted in English and get working code — but the copy-paste cycle was maddening. In 2023, tool use cracked the glass. In 2024-2025, full coding agents put it all together.

The solution? The same one humans discovered thousands of years ago. You do not build a cathedral with one brilliant architect. You build it with a team — specialists who each focus on what they know best, coordinated by someone who holds the big picture.