Skip to content

Instantly share code, notes, and snippets.

@timabell
Created May 11, 2026 21:50
Show Gist options
  • Select an option

  • Save timabell/9e41509f36416f81169d5d7b5fccd888 to your computer and use it in GitHub Desktop.

Select an option

Save timabell/9e41509f36416f81169d5d7b5fccd888 to your computer and use it in GitHub Desktop.
ChatGPT: AI coding history June 2021 - May 2026

The Last Five Years of AI Coding

Five years ago, AI coding tools were mostly autocomplete gimmicks. They could save keystrokes, generate boilerplate, and occasionally surprise you with something useful, but nobody serious thought they were close to replacing real engineering work.

Today, frontier systems can operate inside repositories for hours, coordinate multiple agents, run tests, debug failures, and ship substantial features with limited supervision.

The speed of progress has been absurd.

The important shift wasn’t just “models got smarter”. The entire paradigm changed several times: from autocomplete, to conversational coding, to tool use, to autonomous agents, to full software engineering systems.


June 2021 — Copilot makes AI coding mainstream

GitHub Copilot launch

This was the first moment AI coding genuinely felt useful to ordinary developers.

Copilot could generate whole functions, tests, regexes, CRUD endpoints, and endless glue code directly inside VS Code. It felt magical compared to previous autocomplete systems.

But it was also wildly unreliable:

  • no repo understanding
  • no execution loop
  • no debugging ability
  • no architectural reasoning
  • lots of plausible-looking garbage

Still, it changed developer expectations permanently.


2021–2022 — Codex and AlphaCode prove the idea

OpenAI Codex paper DeepMind AlphaCode

Codex showed LLMs could solve real programming tasks from natural language prompts.

AlphaCode then shocked people by performing competitively in programming contests.

Neither system was practical as an engineering tool yet, but they proved these models could reason about code at a much higher level than most people expected.


Late 2022 — ChatGPT changes the interface

ChatGPT launch

This was arguably bigger than Copilot.

Coding became conversational instead of autocomplete-driven. Suddenly you could ask:

  • “why is this failing?”
  • “refactor this”
  • “write tests”
  • “explain this Kubernetes config”
  • “convert this SQL”

AI became a collaborator rather than a completion engine.

The quality was still rough:

  • hallucinated APIs
  • weak reasoning
  • tiny context windows
  • terrible repo-scale understanding

But this is when AI coding went fully mainstream.


March 2023 — GPT-4 becomes the first serious coding model

GPT‑4 announcement

GPT-4 was the first model many senior engineers found genuinely useful.

It became capable of:

  • non-trivial debugging
  • cleaner abstractions
  • meaningful refactors
  • decent architectural suggestions
  • useful test generation

The conversation shifted from:

“Can AI write code?”

to:

“How much engineering work can it do?”


2023 — The first agent craze

AutoGPT repo

AutoGPT and similar projects introduced the core idea that now dominates the field:

the future is agents, not autocomplete

Most early agents were terrible:

  • infinite loops
  • runaway costs
  • broken memory
  • chaotic outputs

But they introduced the key pattern:

  • plan
  • execute
  • inspect
  • retry
  • use tools
  • continue autonomously

That turned out to matter enormously.


Mid 2023 — Execution loops arrive

ChatGPT Code Interpreter release notes

This was a major conceptual leap.

Models could now:

  1. write code
  2. run it
  3. inspect results
  4. retry automatically

That closed feedback loop massively improved practical usefulness, especially for debugging and data work.


2024 — AI starts acting like a software engineer

SWE-bench paper Devin launch

SWE-bench was important because it moved benchmarks from toy functions to real GitHub issues in real repositories.

Then Devin landed and detonated the industry narrative overnight.

For the first time, people saw a believable demo of an AI system:

  • reading tickets
  • editing repos
  • running tests
  • debugging failures
  • opening PRs

Whether every demo fully matched reality almost didn’t matter. The direction was now obvious.


Mid–Late 2024 — Repo-aware coding explodes

Claude 3.5 Sonnet

This was when coding agents became genuinely useful day to day.

Tools like Cursor, Windsurf, Roo Code, Continue and Aider started combining:

  • long context windows
  • repo indexing
  • terminal access
  • multi-file editing
  • autonomous iteration

“Vibe coding” became a real workflow.

Humans increasingly shifted from typing code to steering and reviewing it.


2025 — Autonomous coding becomes operational

Claude Code OpenAI Codex agent GitHub Copilot coding agent

The workflow changed completely.

Instead of:

“help me write this”

it became:

“go implement this issue”

Agents could now:

  • inspect repositories
  • run terminal commands
  • execute tests
  • patch multiple files
  • debug failures
  • work asynchronously
  • open PR-style outputs

Multi-agent systems also started appearing:

  • planner agents
  • coding agents
  • testing agents
  • review agents

The bottleneck shifted from generation quality to orchestration and verification.


2026 — The current frontier

Steve Yegge on coding agents

At the cutting edge, “autocomplete” is basically obsolete.

Modern systems can:

  • operate for hours
  • coordinate multiple agents
  • manage large repos
  • maintain persistent memory
  • run long debugging loops
  • ship substantial production-quality features

The human role is increasingly:

  • defining goals
  • setting constraints
  • reviewing outputs
  • validating architecture
  • handling edge cases

The remaining hard problems are:

  • reliability
  • verification
  • security
  • architectural coherence
  • subtle regressions

But the direction is now unmistakable.

Five years ago AI coding meant:

“write me a for-loop”

In 2026 it increasingly means:

“own this engineering task end-to-end”

https://chatgpt.com/c/6a01d55a-64ec-83eb-92a9-b47d807b2365

Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment