coding agent
An AI program that independently writes, runs, tests, and fixes code based on a task formulated in everyday language, in multiple steps, until it reaches a working solution or gets stuck.
In plain terms
A coding agent does not just give advice on programming but performs the work itself. You tell it what you want, for example, that 'a coupon code can be redeemed in the webshop.' It examines the existing program, writes the necessary parts, tests them, and if it finds an error, corrects it.
It is like calling in a professional when renovating your apartment, rather than asking for advice at a hardware store. The professional assesses the situation, brings their tools, does the work, and shows you at the end what has been done. You say what you want and you take delivery of the work, but you do not tighten the screws yourself.
Earlier programming assistants only suggested the next line or answered questions in a chat window. The coding agent, however, has access to files and the command line, works in multiple steps, and checks its own results. This allows it to proceed unsupervised on a task for minutes or even hours.
How it works
- It receives the task in everyday language, such as a bug report or a description of a new feature.
- It surveys the terrain: reads the project files, identifies affected parts, and creates a plan for the steps.
- It uses tools: edits files, runs commands, starts tests. Meanwhile, a large language model decides in the background what the next step should be.
- It checks the result. If the program reports an error or the test fails, it reads the error message and tries again. It repeats this loop until the task is completed or it gets stuck.
- At the end, it summarizes what it changed and hands the work over to a human for approval. In most systems, you can set what requires prior permission.
An everyday example
On a small business's appointment booking page, guests do not receive a confirmation email. The developer describes this in one sentence to the coding agent, which finds the faulty part, fixes it, runs the tests, and shows the changes. The developer reviews them and only then activates the fix.
Who is affected
A common misconception
Limits and risks
- It can confidently write faulty or security-vulnerable code, so always review it or have it reviewed by an expert before live use.
- Since it executes commands and modifies files, incorrect settings can cause damage, such as deleting data. Grant only the permissions necessary for the task.
- In large, complex systems and long tasks, it may lose the thread or choose a solution that does not fit the rest of the project.
- Hidden instructions in texts read from external sources, such as web pages or error reports, can mislead (prompt injection).
- Using it costs money, and the expense adds up quickly during long runs.
Similar, but not the same
- code completion assistant – The code completer suggests only the next lines while you type, whereas the coding agent independently carries out a full task in multiple steps.
- vibe coding – Vibe coding is a workflow where a person directs the AI based on intuition without reading the code. This often involves using a coding agent, but agents can also be used for careful, verified work, while vibe coding can be done with a conversational AI.
- chatbot – A general chatbot typically responds with text or a code snippet that you must paste and test in your own project. A coding agent directly accesses the project files and runs the program itself.
Related concepts
AI-agent, generative AI, agentic AI
command-line coding agent, background coding agent, code review agent
vibe coding, tool use, large language model, multi-agent system, prompt injection
A brief history
AI that writes code became widely known in 2021 with GitHub Copilot's line completion feature. At that time, Copilot was based on OpenAI's Codex model. Autonomously working coding agents came into the spotlight in spring 2024 when Cognition introduced a system called Devin. The research agent SWE-agent also appeared that year. In 2025, major AI companies released their own coding agents one after another, such as Anthropic with Claude Code and OpenAI with the new Codex agent, named after the old model.
Where things stand
By the end of 2026, the coding agent is already a daily work tool among software developers and is considered one of the most mature application areas for AI agents. Outside the profession, as well as in large enterprises and strictly regulated environments, its adoption is still ongoing.
In the news
- AI coding agents still need human rework on three-quarters of tasksNov 8, 2026 · Code
- AWS launches Kiro bots to coordinate AI coding agentsNov 5, 2026 · Code
- Meta acquires AI coding startup ConductorNov 4, 2026 · Code
- Cognition launches SWE-2 Flash to make coding agents cheaper and fasterNov 4, 2026 · Code
- AI coding agents are installing MCP servers that will never be patchedNov 3, 2026 · Code
- Coding agents get spending limits as Cognition adds hierarchical budgets to DevinNov 2, 2026 · Code
- Android Studio opens to any AI coding agent, ending Gemini exclusivityNov 2, 2026 · Code
- DeepSeek unveils proactive coding agent prototype to challenge AnthropicOct 31, 2026 · Code
Also known as
coding agent · AI coding agent · AI coding agent · coding agent · software developer agent · software engineering agent · agentic coding · Coding agent
Sources
- Jimenez et al.: SWE-bench: Can Language Models Resolve Real-World GitHub Issues? (arXiv, 2023)
- Yang et al.: SWE-agent: Agent-Computer Interfaces Enable Automated Software Engineering (arXiv, 2024)
This entry was machine-edited and has undergone machine expert review; the “In the news” section updates automatically from Agent Age stories.