Andrej Karpathy coined "vibe coding" in February 2025 - a year later, 41% of all code is AI-generated, agents run multi-hour tasks autonomously, and the developer role has shifted from writing code to orchestrating systems.

Links

• Notes and resources at ocdevel.com/mlg/mla-22
• Try a walking desk - stay healthy & sharp while you learn & code
• Generate a podcast - use my voice to listen to any AI generated content you want

In February 2025, Andrej Karpathy posted a tweet describing how he'd stopped reading diffs, hit "Accept All" on every suggestion, and just copy-pasted error messages back into the chat. He called it "vibe coding" - fully giving in to the vibes and forgetting the code even exists. The post got 4.5 million views. By late 2025, Collins Dictionary named it Word of the Year.

But this wasn't a sudden invention. It was the culmination of a four-year arc that started with GitHub Copilot's line-by-line autocomplete in 2021 and accelerated through GPT-4, 192K+ token context windows, reasoning models, and tool-use architectures. The result: AI shifted from suggesting the next line to autonomously planning, editing, testing, and committing across entire codebases.

The tool landscape has stratified fast

The ecosystem now breaks into three categories:

Terminal-native agents like Claude Code and Gemini CLI give power users direct environment access, scriptability, and Unix-style composability. Claude Code runs on models up to Claude Opus 4.5, supports 200K tokens (1M in beta), and spawns subagents for parallel work. Gemini CLI counters with a 1M-token context window and the most generous free tier in the space - 60 requests/minute, 1,000/day.

IDE-integrated agents like Cursor and Windsurf meet developers where they already work. Cursor hit $1B+ annualized revenue and a $29.3B valuation by going agent-first - its 2.0 release runs up to 8 parallel agents via git worktrees. Windsurf was acquired by Cognition (Devin AI) for $3B.

Cloud-based agents like OpenAI Codex take a different approach entirely - each task spins up an isolated sandbox with your repo, enabling true parallel execution. GPT-5.1-Codex-Max was the first model natively trained for multi-context operation, capable of 24+ hours of independent work.

Open-source pioneers still matter too. Aider (39K GitHub stars) introduced RepoMap for structural code context and now writes 50-88% of its own code. Cline (56K stars) established the human-in-the-loop approval pattern. GPT-Engineer evolved into Lovable, now a $6.6B unicorn.

Three pillars define the emerging stack

MCP (Model Context Protocol) solves the integration problem. Released by Anthropic in November 2024 and now hosted by the Linux Foundation, it's the "USB-C for AI" - a standard protocol replacing N×M custom integrations with N+M implementations. It has 97M monthly SDK downloads and clients across Claude, Cursor, Windsurf, Zed, and VS Code.

Skills turn prompt engineering into reusable packages. They're markdown files that extend agent capabilities through instruction injection - structured recipes telling an agent how to perform specific tasks. They can be shared, version-controlled, and scoped from global to project-level.

Harnesses are the real differentiator. Two agents running the same model differ entirely based on harness quality - the infrastructure governing context bridging, progress tracking, and environment management across sessions. The recommended pattern uses a two-agent architecture: an initializer sets up the environment, and a coding agent makes incremental progress one feature at a time.

Context engineering is the new critical skill

The practical constraint isn't model intelligence - it's what fits in the attention window. The discipline of context engineering has three strategies: reduce (compact older tool calls), offload (save results to filesystem), and isolate (spawn sub-agents for token-heavy subtasks). KV-cache optimization alone delivers 10x cost reduction on repeated context.

What's next

Dario Amodei claimed AI would write 90% of code within 3-6 months of March 2025. Gartner projects 40% of enterprise apps will use AI agents by end of 2026. The near-term trajectory includes repository intelligence (AI understanding code relationships and history, not just lines), production MCP deployments, and agent monitoring with ROI measurement.

The practical takeaway: developers are becoming AI conductors - using agents for boilerplate and rapid prototyping while applying judgment for architecture, direction, and safety. Reviewing AI-generated code effectively requires deeper understanding, not less. The teams winning are those treating infrastructure as lightweight scaffolding around rapidly evolving model capabilities, and expecting to re-architect as models improve monthly.