Agentic coding: How an AI assistant plans, edits, builds and fixes code on its own

Agentic coding is a way of programming in which an AI does more than suggest the next line. It takes a goal, makes a plan, changes several files, runs the project and fixes what breaks. The developer reviews the result instead of typing every step.

The term covers the newest generation of coding assistants, including the ones now arriving inside Apple’s Xcode.

One bug report, handled two ways

Picture a note from a tester: “The settings screen crashes when the username is empty.”

With a classic autocomplete assistant, the developer finds the file, starts typing a fix and accepts or rejects suggested lines. Every decision stays with the person.

With an agent, the developer pastes the note and waits. The agent searches the project, finds the screen, edits the code, builds the app, runs the tests and tries again if something fails. It then presents the changes for review.

The loop: Look, act, check, repeat

Underneath, an agent is a language model placed inside a loop. It reads the task, picks an action, sees the result and decides what to do next.

The actions come from tools the host program provides. Typical ones are searching the code, opening and editing a file, running a build, running tests and reading the error output.

A human developer works the same way: try something, read the error, adjust. The difference is that the model drives the loop and the tools give it real feedback. A failed test is a concrete signal that the last edit was wrong.

Apple describes this in its Xcode 26.3 announcement of February 3, 2026. Agents can plan multistep work from a project’s structure, search documentation, navigate files and change project settings. They can also check their own output by capturing Xcode Previews and running repeated builds and fixes.

Where the idea came from

The loop has a research root. In 2022, Shunyu Yao and colleagues published ReAct, short for reasoning and acting. It showed that a language model does better when it alternates written reasoning with actions, such as querying an outside source, and reads the results before continuing. The authors reported that this reduced the hallucination and error propagation seen with reasoning alone on question-answering tasks.

In 2023, Carlos Jimenez and colleagues released SWE-bench, a test built from real GitHub issues in open-source projects. At publication, the best model they tested, Claude 2, resolved only 1.96% of the issues. The benchmark gave the field a shared measure of progress on realistic software tasks, and later agent systems have been compared against it.

Neither paper is a coding product. Together they supply the two ingredients: a loop that uses feedback, and a way to measure whether it fixes real bugs.

Who offers it today

Apple’s Xcode 26.3 added agentic coding with Anthropic’s Claude Agent and OpenAI’s Codex. Apple also exposed Xcode’s capabilities through the Model Context Protocol (MCP), an open standard, so any compatible agent or tool can connect.

GitHub is following. According to its changelog, a public preview of Copilot inside Xcode’s built-in coding assistant starts with Xcode 27.3 beta. Details are in our news report. The announcement does not say which models or agent features the Xcode integration offers.

Agents also run away from the editor. GitHub’s documentation says Copilot can research a repository, plan changes and implement them in the background, editing files and running tests and linters in a cloud environment. The user reviews the proposed changes before a pull request, which is a request to merge changes into the main code, is created.

Limits worth knowing before trusting one

An agent can be wrong with more confidence and at larger scale than an autocomplete suggestion. GitHub’s own documentation warns that Copilot can produce incorrect or suboptimal code, including code that contains security vulnerabilities, and says to review and test the output before production use.

Several practical limits follow from the design:

• Passing tests is not the same as correct code. An agent that sees green checks may stop, even if the tests miss the real problem.
• Permissions matter. An agent that can run commands can also run the wrong one, so tools usually ask for approval or work in a sandbox.
• Sessions have bounds. GitHub states that a session on GitHub Actions has a maximum execution time of 59 minutes.
• Review still costs time. Reading a large change carefully can take longer than expected.

Outside content is also a risk. If an agent reads text from the web or a repository, that text may carry hidden instructions meant to steer it.

What to watch next

Three things are worth following. First is standard connections: with MCP, one agent can work in several editors, and editors can host several agents. Second is whether vendors publish clear details on cost, such as how agent work counts against a plan’s allowances. Third is measurement: SWE-bench started the habit of testing on real issues, and newer evaluations keep trying to separate genuine fixes from lucky ones.

For developers, the practical skill is shifting from writing every line to describing tasks well, setting up tests, and reviewing changes with care.