The AI Harness: Why Enterprises Are Betting on Platforms, Not Models
There’s a quiet revolution happening in how companies approach AI-assisted software development, and it’s not about who has the smartest model. Personally, I think the real story here is the rise of the AI harness—the layer of infrastructure that sits between developers and large language models (LLMs). What makes this particularly fascinating is that companies like Coinbase, Shopify, and Ramp are all building their own harnesses while still relying on external models like Anthropic’s Claude or OpenAI’s Codex. From my perspective, this isn’t just a technical decision; it’s a strategic one that redefines where the competitive advantage lies in enterprise AI.
The Harness: The New Battleground
One thing that immediately stands out is how these companies are converging on a similar architecture. Instead of trying to outdo frontier models like Anthropic’s or Google’s, they’re focusing on the execution environment—the harness. This layer handles context, permissions, workflow orchestration, and verification. What many people don’t realize is that this isn’t just about making AI work better; it’s about controlling how AI integrates into the unique fabric of each organization.
Take Coinbase’s Forge, Shopify’s River, or Ramp’s Inspect. Each of these tools solves remarkably similar problems but in ways tailored to their specific needs. For example, Shopify’s River participates in one out of every eight merged pull requests, which is impressive. But what’s more interesting, in my opinion, is why: it’s not just the AI agent; it’s the underlying developer platform—the harness—that makes this possible. This raises a deeper question: if the model is just a commodity, what’s the real differentiator?
Why the Harness Matters More Than the Model
If you take a step back and think about it, the harness is where the rubber meets the road. It’s the layer that ensures AI doesn’t just spit out code but does so in a way that aligns with organizational policies, security standards, and cost constraints. A detail that I find especially interesting is how Coinbase reduced its AI spending while processing more tokens. They didn’t achieve this by switching models; they did it by optimizing the harness—intelligent routing, prompt caching, and better visibility into AI usage.
This isn’t just about cost savings, though. What this really suggests is that the harness is becoming the control plane for enterprise AI. It’s where decisions about model selection, workflow integration, and cost optimization are made. Model providers are increasingly becoming interchangeable infrastructure, while the harness is where the strategic value lies.
The Complementary Role of Commercial Tools
Here’s where it gets nuanced: despite building their own harnesses, these companies still rely on commercial tools like Claude Code or Codex. Why? Because, in my opinion, the harness and commercial tools serve different purposes. Internal agents handle asynchronous workflows—think bug fixes or pull requests—while commercial tools dominate interactive sessions where developers work directly in their editors.
What this really highlights is that AI-assisted development isn’t a zero-sum game. It’s not about replacing one tool with another; it’s about creating a symbiotic ecosystem. This duality is something I think many organizations are still wrapping their heads around.
The Platform as the Strategic Asset
The most significant implication of this trend, in my view, is that the platform—not the model—is becoming the strategic asset. A decade ago, companies differentiated themselves through cloud platforms and deployment pipelines. Today, they’re doing the same with AI harnesses.
This shift has broader implications. Model providers will increasingly compete to be the preferred reasoning engine inside platforms they don’t control. If you think about it, this is a power inversion: the platform owners hold the keys, not the model providers. This dynamic is reminiscent of how enterprises built opinionated platforms on top of AWS or Azure without trying to replace them.
The Unpredictable Economics of AI
One thing that’s often overlooked is the unpredictability of AI costs. Walmart and Uber’s recent struggles with AI budgets are a case in point. What many people don’t realize is that autonomous coding workflows can consume up to 1,000 times more tokens than interactive sessions. This makes budgeting a nightmare, and it’s why cost optimization is becoming a core responsibility of the harness, not the model.
From my perspective, this unpredictability is here to stay. Enterprises that own the harness will be better positioned to manage these costs, but it’s not a silver bullet. The real challenge is balancing innovation with fiscal responsibility, and that’s a problem the harness is uniquely equipped to address.
The Future: Platforms Over Models
If there’s one takeaway I’d leave you with, it’s this: the future of enterprise AI isn’t about owning the smartest model; it’s about owning the platform that governs how that model is used. The harness is where context, security, and economics converge, and it’s where companies will differentiate themselves.
Personally, I think we’re only scratching the surface of what’s possible. As common agent runtimes and open protocols mature, the competition will increasingly move above the foundation model. Model providers will become commoditized, while platform owners will hold the strategic high ground.
So, the next time someone asks you whether to build or buy an AI coding assistant, remember: the real question is whether you’re willing to invest in the harness. Because in the world of enterprise AI, that’s where the game is being won.
