You’ve felt it. You’re using Claude Code or Codex, paying premium prices for the smartest AI models on the planet, and the experience feels… mildly shitty. It’s clunky. It’s slow. And it’s draining your wallet faster than a bad habit.
You probably blame the model. You think the AI is just having an off day or hitting a context limit. But you’re wrong. The model is fine. The problem is the bloated, over-engineered wrapper wrapped around it—the “harness.” And big AI companies are charging you a massive premium to suffer through it.
A recent analysis dropped a bombshell: swap out a heavy corporate harness for a minimal one, and you can cut your coding agent costs in half. Without downgrading the model. The harness isn’t a neutral shell; it’s a speed bump disguised as a feature.
You aren’t paying for intelligence anymore; you’re paying a premium to be babysat by a bloated wrapper.
Defenders of the tech giants will immediately cry foul. “But the extra weight is for security and alignment!” they’ll argue. “We need guardrails!” Maybe. But when benchmarks conveniently ignore the cost of these guardrails, optimizing for them doesn’t eliminate risk—it just turns your insecurity into a negative externality. You pay the tax, and you still shoulder the risk.
Look at the trajectory of these tools. Claude Code feels like a vibe-coded project that got out of hand—fucking ridiculous for a $2 trillion company’s flagship product. They keep adding complexity, context windows, and delegation patterns. But the real endgame isn’t more scaffolding.
As the models get smarter, you need to tell them less. The optimal strategy is to let the harness shrink as the model improves. The leading agents are accumulating complexity while the actual endgame is a model that needs almost no scaffolding at all.
When the wrapper costs as much as the engine, you’re no longer buying software—you’re buying a liability disguised as a feature.
If you’re choosing or building coding agents right now, you need to separate model capability from harness overhead. Stop conflating the two, or you will overpay, misattribute performance, and miss the massive shift toward minimal-agent systems.
The companies that win the agent wars won’t be the ones with the most scaffolding. They’ll be the ones brave enough to strip it all away.
The future of AI isn’t building a better babysitter; it’s firing the babysitter entirely.
FAQ
Q: Doesn't the heavy harness provide necessary security and alignment?
A: It's a trade-off, not a guarantee. Big companies use 'security' to justify bloat, but if benchmarks ignore the cost, they're just turning insecurity into a negative externality you pay for.
Q: How do I avoid overpaying for coding agents?
A: Separate the model's raw capability from the harness overhead. Test minimal wrappers with the same models—you might find you're paying double for features you don't need.
Q: What's the real endgame for AI coding tools?
A: The scaffolding disappears. As models get smarter, they need less instruction. The best strategy is to let the harness shrink as the model improves.