How do AI coding agents like Claude Code work?

In detail

A model, a loop and a few tools

A coding agent is a language model wrapped in a program, often called an agent harness, that gives it tools and runs it in a loop. Without tools, the model can only reply with text. With them, it can read a repository, change files and run commands.

Malte Ubl of Vercel recommends the simplest architecture and credits it for how well Claude Code works. Something triggers the agent, it gets input data and a prompt, and it has a set of tools that should be relatively simple. For a coding agent, he lists reading files, listing files, editing files and creating pull requests, “nothing crazy,” just enough for the agent to go figure out what is in the repository (20:48). You describe to the model what each tool is for and what the goal is, and let it work.

Anthropic’s Claude Code documentation, checked October 25, 2026, describes the same shape as three phases: gather context, take action and verify results. Asked to fix failing tests, it says the agent might run the test suite, read the error output, search for the relevant source files, read them, edit them and run the tests again. Each tool result gives it new information for the next step. How a model chooses and calls a tool is covered in how AI agents use tools.

Why it works on code first

Code gives an agent a way to check its own work. A test suite, a compiler or a linter gives it feedback, so the agent can try an edit, see the result and try again without a person judging each step. A passing check covers what that check measures; weak or incomplete tests can still miss a broken change.

The same property helps in training. Ubl says the agreed-upon reason agents are so effective in software engineering, and not yet in other disciplines, is that you can apply reinforcement learning to this valuable, open-ended problem in a way that is harder for work with no source code to run unit tests on. He hedges it: it is a little easier to make agents good at software engineering than at anything else, and he expects the pattern to spread to other fields. The training side is in how reinforcement learning improves AI coding models. Public benchmarks lean on the same check: SWE-bench gives a model a codebase and a real GitHub issue. Its evaluation harness counts a task as fully resolved only when the designated failing tests now pass and the previously passing tests still clear its regression checks.

What fills its context

Everything the agent reads goes into its context window: the conversation, file contents, command output and the project instructions it loads. Claude Code’s documentation says context fills up as you work, that it compacts automatically near the limit, and that instructions from early in a session can get lost. A new session starts with a fresh window and none of the previous conversation, which is why project instruction files such as CLAUDE.md or AGENTS.md exist.

Where people stay in charge

The documentation names two safety mechanisms. Checkpoints let you undo the agent’s file edits, but files changed only through shell commands aren’t tracked, so version control is still the backstop. Permission modes and rules control what the agent may do without asking. Anything that reaches a remote system, such as a database, an API or a deployment, can’t be undone by a checkpoint, so those actions depend on permissions, as with any human-in-the-loop control.

The other check is review. Graphite’s Greg Foster says you don’t want Claude Code, Cursor or similar tools to produce 2,000- or 3,000-line pull requests and hope they pass review and CI. He says Graphite’s data shows that longer changes get fewer review comments per line, and recalls research he thinks included a Google study. That research reference is his recollection, not a study independently checked here. His advice is many small, dependent changes, which he says both people and AI handle better. What happens to review when agents write most of the code is in code review is the new bottleneck.

Disclosure: Episodes 39 and 40 were sponsored by Galileo, where Conor then worked; in episode 40 he points listeners to Galileo’s LinkedIn page at 2:05.