Skip to content
Search
DevelopmentGlossary

Agentic Coding

Where autocomplete suggests the next few lines, agentic coding hands the agent a whole task, such as fixing a bug or adding a feature, and lets it work through the repository with tools: search the code, read files, make edits, run the tests and try again when they fail. The SWE-bench paper defined a common way to measure this, by asking models to resolve real GitHub issues in real repositories.

It differs from vibe coding in who checks the work. In agentic coding the developer writes or approves a clear task (often a spec), the repository's tests decide whether the change works, and a person reviews the diff before it merges. Three things make the biggest difference: an AGENTS.md that tells the agent how to build and test, a good harness with sensible permissions, and tests worth trusting. Qurtoo's articles on system prompts for coding agents, test generation prompts and agentic code review cover each part.

Example

Task: Fix issue #212. Exporting an invoice with no line items crashes the CSV export.
Before you change code:
1. Write a failing test that reproduces the crash.
2. Tell me the cause in two sentences.
Then fix it with the smallest change that makes the test pass.
Run the full test suite and the type check. Show me the diff and the test output. Don't commit.

When to use it

  • Well-defined tasks in a repository with good tests
  • Repetitive changes across many files
  • Investigating a bug, where the agent can run and read code faster than you

When not to use it

  • Repositories with no tests, where nothing checks the agent's work
  • Security-critical code merged without a person's review
  • Vague goals you haven't thought through yet