AI Coding Agent
An AI coding agent is an LLM-driven system that doesn't just suggest code but acts on a codebase autonomously: it reads files, runs tools, edits source, executes commands, and iterates in a loop until a development task is done.
The distinguishing trait of an agent (versus an autocomplete or a chat assistant) is the agent loop: the model is given a goal and a set of tools, then it plans, calls a tool, observes the result, and decides the next step on its own. A typical loop iterates over tools like read file, search, write/edit, run shell command, and run tests, feeding each tool's output back into the model's context so it can self-correct. The loop ends when the model judges the task complete, hits a step limit, or asks the human to intervene.
Most agents share the same primitives: a tool-calling LLM, a context window that holds the relevant code and conversation, some retrieval mechanism to find the right files (grep, embeddings, or a code graph), and an execution sandbox for running commands and verifying changes. Where they differ sharply is in how they *understand* code and how *precisely* they edit it. Many agents operate on raw text — they grep for matches and apply find-and-replace or unified diffs — which is fast to build but brittle: a rename can hit a comment, a string literal, or the wrong scope, and a malformed diff can corrupt a file.
Practical concerns dominate real-world use. Context management matters because code is large and windows are finite; agents prune, summarize, or compact history to stay within budget. Cost and latency add up across a multi-step loop, which is why some agents route different steps to different models. Safety matters because an agent that runs shell commands and rewrites files can do real damage, so checkpoints, diffs, approval gates, and sandboxing are common. Verification — running the typechecker, linter, or test suite inside the loop — is what separates an agent that plausibly edits code from one that reliably ships working code.
The category spans IDE-embedded agents, terminal/CLI agents, and cloud-hosted autonomous agents, and overlaps with the broader idea of agentic coding. The strongest agents are converging on the same lesson: the bottleneck is rarely the model's raw ability and almost always the quality of the code understanding and editing tools you put in its hands.