What is a coding agent? It can take an issue, inspect a repository, run tests and open a pull request

Lesson 033 introduced coding agents. By 2026, it is useful to draw the boundary clearly:

Autocomplete
≠ Chat assistant
≠ Coding agent

Autocomplete predicts nearby code

Classic AI coding assistance looks at the current file and nearby context to suggest the next line or function body.

It is fast and useful, but it may not understand the entire repository or execute anything.

Chat assistants explain and propose

A chat assistant can read supplied code, explain a bug or propose a patch. You may still be responsible for editing files, running tests and opening a pull request.

A coding agent executes a workflow

A coding agent combines a model with tools and an execution loop.

Give it a task such as:

Issue: the login error message disappears too quickly.
Find the cause, fix it, add a regression test, and do not change the API contract.

A capable agent may then:

1. search the repository,
2. locate relevant components,
3. read existing tests,
4. reproduce the bug,
5. edit multiple files,
6. run tests and linting,
7. inspect failures,
8. revise the patch,
9. create a diff,
10. open a pull request (PR).

That is an engineering workflow, not a text-completion task.

Repository context matters

Lesson 027 explained repositories and Git workflows.

A coding agent needs to discover the project’s source layout, test commands, conventions, dependencies, CI rules and protected files. That is why strong agents often spend significant time reading and searching before editing.

Tools are central to agency

Useful tools include file read/write, search, terminal commands, Git, test runners, browsers and GitHub APIs.

This is the same principle from Lesson 020: an agent’s practical capability depends both on the model and on the tools it can operate.

Tests close the loop

Generated code can look plausible and still fail.

A useful coding-agent loop is:

Edit
→ test
→ inspect failure
→ revise
→ test again

An agent that only writes a patch and declares success is missing the verification stage.

A pull request is a safer delivery boundary than direct main edits

A PR makes the change reviewable before it enters the main branch. Humans can inspect the diff, CI can run, and reviewers can request changes.

GitHub’s current coding-agent workflows similarly center on assigning tasks and reviewing the resulting pull requests.

Powerful coding tools require powerful safeguards

If the agent can read secrets, install dependencies, execute shell commands, push code or publish packages, it has meaningful supply-chain and operational power.

Use least privilege, sandboxing, CI, secret scanning, dependency checks and human review. GitHub documents security checks such as CodeQL and secret scanning for agent-created changes, but automated checks are not a substitute for review.

One thing to remember

A coding agent is not simply a better code generator. It can execute a multi-step repository task—read, search, edit, run commands and tests, then return a reviewable diff or PR. Tools, verification and permissions are as important as the model.