Agentic Coding
Also known as:
Agentic software development
Agentic AI coding
Coding agents
Definition
Agentic coding is software work where an AI agent plans steps, edits files, runs commands, and loops until a task looks done, while a person sets the goal, the boundaries, and the review. Autocomplete suggests the next line in the file you have open. A coding agent takes a task, uses tools, and comes back with a diff. You still own the merge. The agent does not join the incident call when that diff breaks production.
Key Takeaways
- An agent is a loop with tools: read, edit, run, look at the result, repeat. Your job is the goal, the limits, and the review.
- Bound the task before it starts. Name the files, the behavior, and the tests. A vague prompt produces a wide diff.
- Two wrong attempts is a signal to take the keyboard. More prompting on a confused thread usually spends tokens and adds code you did not ask for.
- Give the agent the least access that can finish the task. A prompt is not a permission boundary.
How It Works
- You describe a task, often with the files, constraints, and a check that can fail.
- The agent calls tools: search the repo, edit a file, run a test, read the output.
- Each tool result goes back into the conversation, and the model decides the next step.
- The loop stops when the agent claims success, hits a limit, or you interrupt it. Reading the diff is still your step.
Where It Is Used
- Cursor Agent, Claude Code, GitHub Copilot’s coding agent, and similar tools are the everyday form. They differ in models and permissions. The loop is the same shape.
- Teams that get value write the constraints down: which directories the agent may touch, which commands it may run, and which changes need a person.
- Amazon’s 2026 SDE role guideline treats this as normal engineering practice and still scores people on customer and business outcomes, not on how many tokens the loop burned.