Agentic coding definition
Agentic coding is software development where an AI agent does the hands-on work. It reads the codebase, edits files, runs commands and tests, and keeps iterating until the task is done. The developer sets the goal, reviews the result and decides what ships.
In agentic coding you give an AI model a task instead of asking it for a snippet. The agent explores the repository, decides which files to open, changes several of them, runs the build or the test suite, reads the errors and tries again. Claude Code, OpenAI Codex, Cursor's agent mode and GitHub Copilot's coding agent all work this way.
The difference from autocomplete is who holds the loop. With autocomplete you write and the model suggests the next line. With an agent the model writes, and you steer, approve and review.
How is agentic coding different from vibe coding?
Both lean on AI for the typing, so the terms get mixed up. Andrej Karpathy coined vibe coding in early 2025 for building by prompt while ignoring the code itself: you accept whatever the model produces and only check whether the app seems to work. Agentic coding describes how the tool operates and says nothing about how closely anyone looks. You can run an agent and still read every diff, keep the tests passing and throw out changes that don't fit the architecture. Most professional use looks like that.
What does an agent need to do good work?
Context about the project, and a way to check its own changes. In practice that means:
- A project instructions file. Claude Code reads CLAUDE.md, and several other tools read AGENTS.md. It lists the commands, conventions and known traps that a new developer would otherwise learn by breaking something.
- Commands the agent can run to verify itself, such as a type checker, a linter, tests or a local dev server.
- Permission boundaries. Let it read and edit freely, and keep deploys, database writes and anything that touches production behind an approval step.
- Tasks with edges. "Add a glossary listing page that matches this Figma frame" goes far better than "improve the site".
Where does agentic coding fit in website projects?
It's strongest on work that spans many files and follows a known pattern. On a Next.js and Sanity site, adding a new page builder block touches the CMS schema, a GROQ query, the generated types and a React component. An agent can carry a change through all four in one run if the repository documents the pattern, which is why this site's repo has a CLAUDE.md that spells it out step by step.
It's weak at taste. Spacing, type and motion still need someone looking at the rendered page on a real phone, because an agent will call a layout finished once it compiles. Review is the other cost. An agent can produce a 600-line diff in a few minutes, and reading it properly can take longer than the agent took to write it. If nobody reads it, you're back to vibe coding with extra steps.
Related terms
- Vibe codingVibe coding is building software by describing what you want to an AI model and accepting the code it writes without reading it closely. You judge the result by running it and paste any errors back into the chat. Andrej Karpathy coined the term in February 2025.
- AI agentAn AI agent is software that uses a language model to work toward a goal on its own, step by step. It decides what to do next, calls tools such as search, APIs or a browser, checks the result and keeps going until the task is done or it needs a human.
- Agentic workflowAn agentic workflow is a multi-step process where a language model handles some of the steps, such as drafting, classifying, calling tools or checking output. The order of steps is fixed in code. That separates it from an AI agent, which plans its own steps.