A coding agent is an AI agent designed to perform software engineering tasks by interacting with a codebase and its development environment. Agentic coding is the workflow in which such an agent plans changes, edits files, runs commands, interprets failures, and iterates toward a verified result.
Unlike a conversational coding assistant that only suggests snippets, a coding agent can operate across the engineering loop:
- inspect repositories, documentation, and version-control history;
- search for relevant symbols and dependencies;
- modify multiple files;
- run formatters, compilers, tests, and static analysis;
- use browsers or other tools to verify behavior;
- diagnose failures and revise the implementation; and
- produce a patch or commit for human review.
Coding agents commonly combine a reasoning model with tool calling for shell commands, file operations, code search, and external documentation. Their effectiveness depends heavily on context engineering, because repository instructions, architecture, test output, and local conventions must remain available without overwhelming the model.
This term overlaps with vibe coding, but the practices are not equivalent. Vibe coding emphasizes generating software from high-level natural-language intent, often with limited attention to implementation details. Agentic coding describes the autonomous execution pattern and can include rigorous planning, code review, testing, and validation.
Coding agents introduce operational risks because they can execute code and modify persistent artifacts. Production environments should use sandboxing, scoped credentials, network restrictions, command approvals, isolated branches or worktrees, and human review before deployment. Generated tests should not be the only evidence of correctness, since an agent may encode the same misunderstanding in both implementation and test.
Evaluation should measure end-to-end task completion, regression rate, maintainability, security, and the amount of human correction required. Benchmarks based only on generating isolated functions do not capture the long-horizon behavior of an engineering agent.
As models improve at sustained tool use, coding agents are becoming a prominent example of agentic workflows applied to professional knowledge work.
The LLM Knowledge Base is a collection of bite-sized explanations for commonly used terms and abbreviations related to Large Language Models and Generative AI.
It's an educational resource that helps you stay up-to-date with the latest developments in AI research and its applications.