How fast is Claude Code growing? Anthropic CEO Dario Amodei gave the number: 80x.
They planned infrastructure for 10x annual growth. They got 80x.
That's why Claude Code users have been hitting compute constraints and usage limits over recent weeks.
The Root Cause
It's not just more users. It's how they're using it.
A year ago, most people used one agent, one terminal window, for simple code tasks. Now, users run six terminal tabs simultaneously, dozens of subagents, complex agent workflows.
Cat Wu gave an example: some users installed plugins that made Claude launch 100 subagents in the background. They didn't even realize it — equivalent to running 100 Claude Codes at once.
This explains why "usage limits" became a user complaint. Anthropic isn't arbitrarily restricting you — some people are just using it massively.
Anthropic's Response
Two things.
First, doubled usage limits for Pro and Max plans. Direct capacity release, but a band-aid solution.
Second, increased transparency. When your cache breaks, Claude Code notifies you "/clear to start a new session — it'll be cheaper." Running /usage shows which sessions consumed heavy tokens.
The subtext: we won't save tokens for you, but we'll show you where they went. You decide.
The "Lean Harness" Philosophy
The most interesting part of the interview.
Wu proposed: as model capabilities improve, Claude Code's tool interface should get simpler, not more complex.
She calls it "lean harness" — a lightweight framework that doesn't over-constrain the model.
Specific example: as models get smarter, the team removes parts of the system prompt, reduces tool descriptions. Not because they're unnecessary — because the model already "understands."
Wu even predicted a possible endgame: "Maybe we all collapse back to a simple text box because the model is always right, so you don't need to follow every step."
This is a bold bet. It means much of what Claude Code does now — multi-agent management, complex toolchains, IDE integration — is essentially transitional scaffolding.
My Take
Claude Code's product philosophy is interesting — it admits it might be a transitional form.
Most software products don't say "this feature we built might not be needed later." But Claude Code's team explicitly said it: the current multi-agent management and complex toolchains exist because models aren't smart enough yet. When they are, these things will disappear.
This honesty is either extreme confidence (model capability will really improve at the expected pace) or risk management (if models don't improve fast enough, there's a rationale for why product form needs to change).
For users, the signal is: don't over-invest in specific Claude Code workflows. The tool itself may fundamentally change in 6-12 months.
But that doesn't mean Claude Code isn't worth using now. It's still one of the best AI coding tools available. Just know that its form is fluid.
Primary source: Ars Technica interview with Cat Wu