A year ago, "AI coding" meant autocomplete with vibes. Today it means a process that reads your GitHub issue and opens a PR.
Anthropic's Model Context Protocol crossed 97 million installs in March 2026, and every major AI provider now ships MCP-compatible tooling. The category has fractured into clear archetypes: terminal agents (Claude Code, Codex CLI), AI-native IDEs (Cursor, Windsurf), cloud-hosted autonomous engineers (Devin, GitHub Copilot Coding Agent), and open-source frameworks where you bring your own model. Total market: $12.8B in 2026, up from $5.1B in 2024.
The defining capability shift this year is long-running autonomous workflows. Agents are no longer single-prompt. They execute loops — for minutes, hours, sometimes overnight — reading the repo, running tests, retrying.
§ 02 A note from someone who lives in this
I ship most of my code with an agent in the loop now. Two non-obvious lessons from the last six months:
- The bottleneck is no longer code generation. It's the eval — knowing whether what the agent produced is correct and shippable.
- The unit of trust is the sandbox, not the model. Whether the agent is "good" matters less than whether you can give it root inside an isolated environment and accept that it might break something there.
The 2024 question was "is the agent smart enough?" The 2026 question is "what container did you put it in?"