Claude Code: the agent that codes, not an autocomplete
80% of Anthropic's production code is written by its own AI. What Claude Code actually does, why it isn't Copilot, and what it takes to get value from it.
By Nacim Moudjeb6 min2
80% of their code is written by their AI
In May 2026, Anthropic dropped a number worth pausing on: more than 80% of the code that ships to production at their company is written by Claude, their own AI. A year earlier, it was a few percent.
Ask the question for your own company. If the firm that builds the model lets AI write eight lines out of ten… what does that mean for your twelve-person engineering team?
Before answering, you need to understand what Claude Code actually is. Because most people file it under the wrong label.
It's not a Copilot. It's an agent.
You know the GitHub Copilot model: you type a line, it completes it. Handy, but you do everything else — open the files, run the tests, fix, repeat. The AI suggests; you execute.
Claude Code works the other way around. You give it a task — "fix this bug", "migrate this module to the new version", "write the missing tests" — and it carries it end to end:
it reads the project (up to one million tokens of context — a very large codebase at once);
it acts: edits multiple files, runs commands, executes the tests;
it verifies: looks at what breaks, traces the cause, fixes, repeats.
It runs that "read → act → verify" loop on its own, without you approving every micro-step. That's what an agent is: it doesn't string text together, it completes a multi-step task. The same logic as our AI agents, applied to code, in your developers' terminal.
What it looks like in practice
Concrete numbers. At Rakuten, the average time to ship a new feature dropped from 24 working days to 5. In one striking test, Claude implemented a method inside a 12.5-million-line codebase in 7 hours of autonomous work, in a single pass, with 99.9% numerical accuracy versus the reference.
And the use goes well beyond pure code. Claude Code plugs into your internal tools via an open standard (MCP): read your docs, update your tickets, query your databases, analyse your logs. "Ping me on Slack if you spot an anomaly in these logs" is a valid command.
And it's not a lab toy. Six months after its public launch, Claude Code passed a billion dollars in annualised revenue, and companies like Netflix, Spotify, KPMG, L'Oréal and Salesforce plugged it into their code. The open standard it uses to connect to your tools — MCP — already counts 10,000+ public connectors and has been handed to a Linux foundation. When players like that line up this fast, it isn't a fad.
The part nobody tells you (and that matters)
Now, the honesty — because the AI hype deserves better than yet another "this changes everything".
That "80% of the code" figure, Anthropic qualifies itself. When they say their engineers merge "8× more code per day", they add in the same breath that the number "is almost certainly an overstatement of the true productivity gain": counting lines measures quantity, not value. Their own median estimate sits closer to 4×. And more than half their employees can fully delegate only 0 to 20% of their work to AI.
Translation: Claude Code is not a developer you take off payroll. It's a supercharged but supervised collaborator. Someone stays in control, arbitrates, validates. The gain is huge; it isn't magic.
For a business owner, the real questions
Two objections always come up, and they're fair.
"Does my code train their AI?" No, not under commercial terms: Anthropic commits to not training its models on the code and prompts sent through Claude Code in professional use (and a "zero retention" mode exists for enterprise). Caveat: that's true for the pro plans, not free personal accounts. The personal-use vs business-use line is real — exactly the kind of thing you settle before rolling it out.
"What about security?" By default, the agent is read-only. Any sensitive action — editing a file, running a command — needs your go-ahead. It only writes inside the project folder, nowhere else. And at company level, you lock it down: which commands are allowed, which tools are connected, which models are permitted. You keep control.
The €10,000 mistake: deploying it alone
Here's the opinion we'll stand behind: a tool this powerful does nothing on its own.
Installing Claude Code and leaving your devs with a YouTube tutorial is like buying a Formula 1 car to do the grocery run. The engine is there; nobody can drive it.
The real gain comes from three things:
training the team to delegate to an agent — a reflex that doesn't come naturally;
coding it custom skills (your conventions, your processes, your internal APIs) so it works like someone who knows the house;
plugging it into your tools (your databases, your tickets, your docs) instead of letting it code in a vacuum.
Claude Code isn't a smarter autocomplete. It's an agent that reads, executes and verifies whole tasks — under control, without training its models on your code. The tech is mature, the proof is in the numbers. What separates "we installed something" from "we saved weeks" is the setup: training, skills in your image, connection to your tools.
Want to know where Claude Code would save your team time? Book a free audit — let's look at your stack together.