Tech, ExplainedCloud platforms and developer tools
Concept

AI coding agents

Writes the code, runs it, tries again.

HOW IT WORKSA person gives theagent a task, such asfixing a failing testThe agent reads therelevant code andwrites a changeIt runs the code orthe tests in a sandboxIt reads theerrors and changesthe code againWhen the result works,a person reviews itRunning the code shows the agent its own mistakes, which a model that only writes code cannot see.HOW IT WORKSA persongives theagent a task,such asfixing afailing testThe agentreads therelevant codeand writesa changeIt runs thecode or thetests ina sandboxIt reads theerrors andchanges thecode againWhen theresult works,a personreviews itRunning the code shows the agent its own mistakes, which a model that only writes code cannot see.HOW IT WORKSA person gives the agent a task, such asfixing a failing test↓The agent reads the relevant code andwrites a change↓It runs the code orthe tests in a sandbox↓It reads the errors andchanges the code again↓When the result works,a person reviews itRunning the code shows the agent its own mistakes, which amodel that only writes code cannot see.
An agent is given a task, writes code, runs it in a sandbox, reads the errors and changes the code again, until a working result goes to a person for review.
opens a page inside this one a page in the same scene a page in another scene a concept page a label only, with no page

AI coding agents are AI programs that are given a programming task, and then write the code, run it, read what happened and try again until the result works. The code can be a change to an app, a database query or a test. Older AI coding assistants only suggested code, and a person had to run it and judge it. An agent runs the code itself, so it can see its own mistakes and correct them.

Here is how it works in general. A person gives the agent a task, such as "fix the wrong total shown on the bill". The agent is an AI agent: it runs on a language model, and it has tools. It reads the relevant files, writes a change and then runs the code or the tests, usually in a sandbox, a separate space where mistakes cannot harm real systems. It reads the errors or the results. If something failed, it changes the code and runs it again. When the result works, it hands the change to a person, often as a pull request, a proposed change that others review before it joins the main code. The AI coding agent product page shows this in a software team.

This loop is useful, but an agent that can run code can also do damage. It might delete files, use a password it was not meant to use, or add a library with a known flaw. So companies run agents with narrow permissions, as described in AI agent guardrails. A person still reviews what the agent produced.

Take, for example, a sales head who types a question about last month's returns. In asking the data in plain language, an agent writes the database query, runs it, fixes the query if it fails and then explains the answer.

Consider a team designing a chip. With AI for chip design, an agent writes verification tests, runs them on the design and reports which ones fail.

Say a security team wants its website tested before attackers try it. Automated red teaming uses an agent that writes attack steps, tries them and changes its plan based on what happens.

How is an AI coding agent different from an AI coding assistant? An assistant writes code when asked and stops. A person must copy it, run it and decide whether it works. An agent runs its own code and checks the result, so it can finish a whole task. This also means it needs clearer limits and a careful review.

Appears in

Networking, infrastructure and security
Cloud platforms and developer tools
Data platforms and AI
Semiconductors and devices
Tech, Explained is a part of ingrid.fyi. Every page explains one piece of technology in plain words.