Agentic coding
Agentic coding is a way of working where an AI agent runs a self-directed loop — reading code, calling tools, editing files, and checking the result — to complete a software task from a high-level prompt, instead of just suggesting one snippet at a time.
The defining trait of agentic coding is the loop. A traditional code assistant takes a prompt and returns text; an agentic system takes a goal and decides for itself which actions to take to reach it. On each turn the model receives the current state, chooses a tool call (read a file, run a test, grep, edit, execute a shell command), observes the result, and feeds that observation back into the next decision. The loop continues until the task is done or a stopping condition is hit. This is what people mean when they say an agent "works on its own" — the autonomy comes from the model controlling its own action sequence rather than a human stepping it through each command.
Three components make the loop work: a reasoning model that plans and decides, a set of tools that give it real capabilities in your environment, and a feedback channel so it can see the consequences of its actions and correct course. The feedback step is what separates an agent from a script. When a test fails or a typecheck errors, a well-built agent reads the error and tries again, the same way a developer would. Quality of tools matters as much as quality of the model: an agent that can only do text find-and-replace will make different mistakes than one that can rename a symbol through a language server.
In practice, agentic coding spans a spectrum of autonomy. At one end, the human approves every action; at the other, the agent runs end to end and reports back. Most real-world setups sit in between, with guardrails like permission prompts, checkpoints, and step limits. The hard problems are not "can the model write code" but the surrounding engineering: keeping the context window focused on what's relevant, avoiding runaway tool loops, recovering from bad edits, and controlling cost as the loop racks up tokens across many turns. Those constraints are why agent design is mostly about context management, tool precision, and safety rails, not just prompt wording.