Recently, some of Anthropic's Claude Code leaked online. A lot of teams got their hands on it, dug into the details, and some even used it as a reference to build their own code agents.
The key takeaway is this: the "magic" isn't the AI itself. It's the processes and system design around it.
What did the leaked code show?
From the outside, Claude Code looks like this incredibly smart, end-to-end system that can handle everything on its own. But when you look at the leaked ~500k lines of code, a different picture emerges. The actual "intelligence" layer that makes decisions is less than 5% of the system. The vast majority is just solid, traditional software engineering.
The AI isn't an independent engine. It's more like a guiding layer. It suggests what should be done, but the parts that actually execute tasks, handle errors, ensure safety, and run the process end-to-end are all part of the system.
Harness — the layer that makes AI productionizable
The real differentiator is what's often called the "harness" — the layer that keeps the AI under control and makes it usable in production.
- Context management: A multi-layered structure that constantly summarizes and organizes conversations so the AI doesn't lose track.
- Safety: Multiple control mechanisms that block risky or incorrect actions early on.
- Tooling: 50+ integrated tools. The AI gives direction, but the real work is done by these tools.
The conclusion is pretty clear: the system looks "smart" because the engineering around it is strong. The model alone doesn't create that effect.
Where is the real competition shifting?
This leads to a bigger shift. The question "Which model is the smartest?" is going to matter less and less as models converge.
The real competition is moving toward:
- Who can orchestrate AI better
- Who can build more robust systems
- Who can manage errors more effectively
In short, it's not about building AI. It's about building systems that work with AI. And the real value sits in that invisible, underlying layer.
