Agentic Coding.
AI-assisted software development where the AI operates autonomously across multiple steps — reading files, executing code, observing results, and iterating — with minimal human intervention per action.
Agentic coding extends AI code generation from single-response interactions into autonomous multi-step workflows. An agentic coding system can receive a high-level task ('add user authentication to this Express app'), plan the necessary changes, read relevant files, write code, run tests, observe failures, and iterate until the tests pass — all without a human approving each intermediate step.
The key components of an agentic coding system are: a capable code-writing LLM, tool access (file read/write, terminal execution, web search), and a planning or orchestration layer that decides what to do next based on intermediate results. Tools like Cursor Agent, Claude Code, Devin, and Aider implement agentic coding with different degrees of autonomy and human oversight.
The tradeoff in agentic coding is control vs. speed. More autonomous agents complete tasks faster but may make architectural decisions the developer wouldn't have chosen, generate code that's hard to review as a coherent unit, or go down unproductive paths. Effective practitioners typically keep humans in a supervisory role — setting goals, reviewing plans, and checking intermediate results — rather than running agents fully autonomously.