The first generation of AI coding tools — GitHub Copilot, Tabnine, Amazon CodeWhisperer — worked as smart autocomplete. They predicted the next line or the next function based on surrounding context. Useful, but fundamentally reactive: the developer typed and the tool suggested. The tools shipping in 2025 and 2026 work differently. They plan, execute, test, and iterate — without the developer touching every intermediate step.
Cursor, the AI-first IDE built on a VS Code fork, has become the reference point for this shift. Its Composer feature, and the more recent Composer Agent mode, allow developers to describe a task in natural language — refactor this module to use the new API, add integration tests for the payment flow, migrate this codebase from Python 2 to 3 — and have the agent execute across multiple files, run tests, observe results, and continue. Cursor reported crossing one billion dollars in annual recurring revenue in 2025, a growth rate that reflects how quickly developer teams have adopted the agentic workflow.
GitHub Copilot Workspace, Microsoft's response to the agentic challenge, operates at an even higher level of abstraction: it starts from a GitHub issue, plans a set of file changes to resolve it, and presents a diff for developer review before any code is written. The developer role shifts from writing code to reviewing plans and approving changes. For experienced developers, this is a multiplier. For less experienced developers, it is a significant risk — the agent's plan can be wrong in ways that are non-obvious to review.
Anthropic's Claude Code, launched as a CLI tool, takes a different design philosophy: it operates directly in the terminal against the actual filesystem, runs real shell commands, and treats the developer's environment as the execution context. Rather than abstracting away the development environment, it operates within it. This design makes it transparent and powerful for developers comfortable with the command line, but demanding for those who are not.
Devin, from Cognition AI, represents the most autonomous end of the spectrum: a fully agentic software engineer that can be assigned tasks via Slack and return completed pull requests without developer involvement in the execution loop. Early benchmarks were impressive; production deployments have revealed that the failure modes of autonomous agents — incorrect assumptions, incomplete error handling, security oversights — require careful human review of outputs.
The practical shift is already visible in how engineering teams are structured. Senior engineers are increasingly spending time reviewing agent-generated code rather than writing code themselves. The volume of code being produced has increased dramatically; the bottleneck has shifted to code review and system design quality. Teams that have adapted to this are shipping faster. Teams that have not adapted are drowning in agent-generated pull requests that need as much review time as hand-written code.