Agentic coding

Core idea

Agentic coding goes beyond returning a code snippet or making one requested edit. Given a goal, an AI agent chooses tools to inspect, edit, run, observe, and adapt across multiple steps. An AI coding agent is the system that uses tools; agentic coding is the workflow in which that system repeats exploration, modification, execution, and verification. Human approval is a safety policy for this course rather than part of the definition.

In plain language

Instead of copying code from a web chat into an editor, you can ask, “Create a screen where students solve a problem and check their answer.” The agent locates relevant files, edits them, runs the project, and responds to what actually happens.

Typical flow

1. A person explains the goal and the allowed scope.
2. The agent inspects the relevant files and current state.
3. The agent edits files, runs the program or checks, and interprets the result.
4. It adjusts and repeats as needed, then the person inspects the result and evidence and decides what happens next.

What remains human work

Tool access does not guarantee a correct result. People remain responsible for instructional purpose, private information, publication, and other consequential decisions. Destructive actions, external releases, and paid operations need appropriate approval and verification.

Professional terminology

Agentic coding connects to AI coding agent, tool use, agent loop, and AI-assisted software development. Vibe coding informally emphasizes expressing intent with limited attention to implementation details; it may or may not use an agentic loop. Multi-agent work is an optional extension for genuinely separate assignments, not a requirement. Larger work also depends on context engineering and harness engineering.

Role in this blog

The OpenCode walkthroughs show the sequence from a request to file inspection, edits, execution, and human review. Beginners start by explaining goals and judging outcomes rather than writing every line themselves.