Skip to main content
Zubnet AI › Learn › Wiki › Coding Agents
Using AI

Coding Agents

Also known as: Agentic Coding, AI Coding Agents
AI systems that carry out programming tasks end to end by reading the codebase, editing files, running commands and tests, and iterating on the results.

Why it matters

Software is an unusually good fit for agents. Code is text, the tools (shell, compiler, tests, version control) are already scriptable, and results can be checked automatically: the tests pass or they don't. That feedback loop lets an agent correct itself in a way that's hard in most other fields.

The effect on how software gets built is large. Developers increasingly act as reviewers and directors of agent work instead of typing every line, and people who aren't programmers can build working applications.

Deep Dive

A coding agent is given a goal, like "fix this failing test" or "add CSV export to the reports page", and works toward it on its own. It searches the repository, reads the relevant files, makes edits, runs the build and the tests, reads the errors and tries again. This is different from autocomplete or a chat assistant, where the model suggests code and the human does everything else.

Coding agents became practical in 2024 and mainstream in 2025, as models got reliable enough at tool use to work for many steps without drifting. They are now one of the most widely used and commercially important applications of LLMs.

The Main Forms

  • Terminal agents: Claude Code (previewed in February 2025, generally available that May), OpenAI's Codex CLI, Gemini CLI and the open-source Aider run in a terminal with direct access to your project.
  • IDE agents: Cursor, Windsurf, GitHub Copilot's agent mode and Cline work inside the editor.
  • Cloud agents: Devin (announced in March 2024), Codex in the cloud, GitHub's Copilot coding agent and Google's Jules work asynchronously in a remote sandbox and come back with a pull request.
  • App builders: Lovable, Bolt, Replit and v0 turn a description into a running web app, and are aimed at people who may never look at the code.

How They Work

Under the hood is a model inside an agent harness: a loop that lets the model call tools to search, read, edit and execute. Most agents also read a project instruction file, such as AGENTS.md or CLAUDE.md, that explains how to build, test and follow the project's conventions. MCP servers and skills extend them with outside tools and procedures. A permission system or sandbox controls what they can run without asking.

Measuring Them

SWE-bench, built from real GitHub issues, is the best-known benchmark. Top scores went from under 2% when it launched in 2023 to well over 70% on its Verified subset by 2025. Terminal-Bench tests harder command-line tasks. Another view is task length: the research group METR has found that the length of tasks agents can complete has been doubling roughly every seven months, and faster in recent years.

Working Well With One

  • Give it a way to check itself. Tests, a type checker, a linter or a screenshot turn guesswork into a feedback loop.
  • Be specific. Name the files, the expected behavior and the constraints. Ask for a plan first on larger tasks.
  • Keep tasks scoped. One clear change per session works better than a sprawling request, and a fresh context beats a long, cluttered one.
  • Review the diff. You are still responsible for what gets merged.

Risks

Agents sometimes satisfy the letter of a task instead of its intent: editing a test so it passes, or hard-coding an expected value. They can produce code that works but duplicates existing logic or ignores the project's design, which adds up to maintenance debt. They are exposed to prompt injection through web pages, issues and dependencies, and they often run with access to secrets, so sandboxing and limited credentials matter. Costs can also climb quickly on long tasks.

← All Terms
ESC