Skip to content

What is agentic coding?

Mikhail Bodnarchuk

Definition

Agentic coding is software development in which an AI coding agent works in a loop: it reads the codebase, plans a change, edits files, runs commands and tests, reads the output and fixes its own mistakes until the task is done. The developer sets the goal and the permissions, then reviews the diff before it merges.

How does agentic coding work?

A coding agent is a language model wrapped in a harness that gives it tools. Anthropic's docs for Claude Code split the work into three phases that repeat: gather context, take action, verify results (How Claude Code works). The model decides each next step from what the last tool call returned.

For a prompt like "fix the failing tests", the same docs list the steps the agent may take:

  1. Run the test suite to see what fails.
  2. Read the error output.
  3. Search for the source files involved.
  4. Read those files.
  5. Edit them.
  6. Run the tests again to check the fix.

The tools are what make it agentic. Without them the model can only answer in text. With them it reads and edits files, runs shell commands and git, searches the web, and calls external services through MCP servers. You can interrupt at any point and steer it.

How is agentic coding different from autocomplete and chat?

The line is drawn by who runs the code and who checks the result.

AutocompleteChat assistantAgentic coding
ScopeThe current line or functionThe snippet you pasteThe whole repository
Runs commandsNoNoYes: builds, tests, git
Checks its own workNoNoYes, by running tests and reading errors
OutputInline suggestionText you copyA diff, a commit or a pull request
Where you reviewAs you typeBefore you pasteOn the finished diff

This page covers the mechanics of the tool. For the wider practice of supervising agents across specs, code, tests and review, read our guide to agentic engineering.

Where does a coding agent run?

Where the agent runs decides what it can touch.

  • On your machine. Claude Code runs locally by default with access to your files, terminal and git state. It can also run in Anthropic-managed cloud VMs or self-hosted environments (Anthropic).
  • In a cloud sandbox or CI runner. GitHub Copilot can plan changes and implement them in the background, editing files and running tests and linters in a cloud environment instead of your IDE. Its environments are a cloud sandbox, a GitHub Actions runner or a remote custom environment (GitHub Docs).

In GitHub's setup, Copilot works on one branch at a time and opens one pull request per task. That keeps each change small enough to review.

What does a safe agentic coding setup look like?

Two files do most of the work. The first tells the agent how to build and verify. Claude Code reads a CLAUDE.md file every session, and it can read an AGENTS.md file that other coding agents use instead (Anthropic).

# AGENTS.md

## Commands
- Install: npm ci
- Test (run after every change): npm test
- Lint: npm run lint

## Rules
- A task is done only when tests and lint pass.
- Never edit files in migrations/ that are already merged.
- Do not add dependencies without asking.

The second limits what the agent may run without asking. This .claude/settings.json uses Claude Code's rule syntax. Rules are evaluated deny first, then ask, then allow, and an allow rule cannot carve an exception out of a deny rule (Claude Code permissions).

{
  "permissions": {
    "allow": ["Bash(npm test *)", "Bash(npm run lint)", "Bash(git diff *)"],
    "ask": ["Bash(npm install *)"],
    "deny": ["Bash(git push *)", "Read(./.env)", "Read(./secrets/**)"]
  }
}

The same docs warn that pattern rules have limits: a push written another way, such as git -C . push, is not matched by Bash(git push *). Treat the deny list as one layer, and keep branch protection on the remote.

How many developers use coding agents?

Use is now mainstream, and mostly supervised.

  • In the Stack Overflow 2026 Developer Survey, coding assistants and agents were the top use of AI at 66%, ahead of general chatbots at 63%. Of respondents who use coding assistants or agents, 73% use them daily. More than 30,000 people took the survey, according to the results announcement of 2026-10-06.
  • In a Stack Overflow pulse survey of 1,100 developers and professionals (published 2026-05-27), agent use rose from 31% to 59%. Still, 63% rarely or never let agents run fully on autopilot, and 60% block agents from making unapproved system changes.

The pattern behind those numbers: teams use agents daily but keep a person on the merge button. The AI in the SDLC trends post covers what that does to productivity and code security.

When does agentic coding work, and when does it not?

It works best when the agent can prove its own result:

  • A failing test or a bug with a reproduction. The test tells the agent when it is done.
  • Mechanical changes across many files. Renames, API migrations, dependency upgrades with a green build as the target.
  • Missing tests for existing code. The current behavior is the spec.

It works poorly without a check to run:

  • No tests, or tests that take an hour. The loop has nothing fast to verify against, so the agent guesses.
  • Vague goals. "Make it better" produces a large diff that is hard to review.
  • Security-sensitive code shipped without review. GitHub states that Copilot can produce code with security vulnerabilities and asks you to review and test output before production (GitHub Docs).

What should you check before relying on it in production?

  1. One command verifies a change. Tests and lint run in minutes, locally and in CI.
  2. Permissions are written down. Push, deploy and secret files are denied or set to ask.
  3. The agent works on a branch. Every change reaches main through a pull request that a person approves.
  4. Connected tools are reviewed. Each MCP server adds actions the agent can take, so list them and scope their credentials.
  5. You measure the outcome. Track review time, reverted changes and escaped defects, not only how much code the agent wrote.

SDCLabs builds these setups for client teams: agents draft code and tests, senior engineers review what ships, and automated tests check every change. See AI-enabled engineering.

Sources

  1. How Claude Code works, Anthropic (Claude Code docs), 2026-11
  2. Configure permissions, Anthropic (Claude Code docs), 2026-11
  3. GitHub Copilot on GitHub.com and GitHub Mobile, GitHub Docs, 2026-11
  4. Stack Overflow Developer Survey 2026: AI, Stack Overflow, 2026-10
  5. The results of the 2026 Developer Survey are here, Stack Overflow, 2026-10-06
  6. Agents on a leash: Agentic AI remains mostly single-agent and monitored at work, Stack Overflow, 2026-05-27
  7. Vibe coding (slang entry), Merriam-Webster, 2026-06-04

Frequently asked questions

What is the difference between vibe coding and agentic coding?

Merriam-Webster describes vibe coding as telling an AI what you want and letting it build the product, where the person does not need to understand how the code works (Merriam-Webster, updated 2026-06-04). Agentic coding describes how the tool works, not how closely you read the result: you can run a coding agent and still review every diff.

How do I get started with agentic coding?

Pick a repository with a test suite that runs in one command, and give the agent a small, checkable task such as fixing one failing test. Write the build and test commands into a CLAUDE.md or AGENTS.md file, and allow only those commands before you widen the scope.

What is the best AI for agentic coding?

There is no single best tool, so compare them on your own repository and tasks. By usage, Claude Code (66%) and GitHub Copilot (59%) lead the coding agents in the Stack Overflow 2026 Developer Survey.

Does agentic coding replace code review?

No. GitHub's documentation for Copilot says its session logs show the work and tools used but do not replace your own review and testing (GitHub Docs). Treat an agent's pull request like one from a new teammate.

Building with AI agents? Talk to an engineer.