There's a lot of noise about AI writing code. Most of it misses the point. The interesting shift isn't autocomplete — it's the agent: a tool that can read a whole repository, run commands, edit files, run the tests, read the failures, and try again. Claude Code is that, and it has quietly become part of how we work.
What it actually is
Claude Code is Anthropic's agentic coding tool. It runs in your terminal (and IDEs, and CI), with access to your codebase and the shell. You give it a goal in plain language; it explores the code, makes changes across many files, runs the build and tests, and iterates until the task is done — narrating what it's doing as it goes.
The difference from a chat window is that it closes the loop. It doesn't hand you a snippet to paste; it makes the change, checks it, and fixes what broke.
Where it earns its place
We don't point it at greenfield toys. It earns its keep on the unglamorous work that eats senior time:
- Migrations and upgrades — bumping a framework across hundreds of files, where the change is mechanical but the volume is brutal.
- Test backfill — writing the characterization tests nobody got around to, against code that already exists.
- Codebase archaeology — "where does this value actually get set?" answered in seconds across a million lines.
- Glue and scaffolding — the tenth CRUD endpoint that follows the same pattern as the other nine.
The win isn't that it writes code. It's that it reads your code first, then writes code that matches it.
The discipline that makes it safe
An agent that can run commands is powerful and, used carelessly, dangerous. The practices that keep it boring:
- Work on a branch, review every diff. The agent proposes; a human merges. No exceptions.
- Let the tests be the guardrail. A strong test suite turns the agent loose safely — it can't quietly break what's covered.
- Give it context, not just instructions. A short
CLAUDE.mddescribing conventions and gotchas makes its output read like the rest of the codebase. - Keep tasks scoped. "Refactor this module and keep the tests green" beats "improve the app."
What it doesn't do
It doesn't replace judgment. Architecture, trade-offs, what to build and what to leave alone — that's still the engineer's job. Claude Code makes a good engineer faster; it doesn't make a non-engineer a good one. The teams getting value treat it as a very fast, very literal junior who never tires and always reads the docs.
That's how we use it on client work — including parts of this very site. If you want help wiring agentic tooling into your own engineering workflow safely, let's talk.