The AI Harness: Why Enterprises Are Building Their Own Coding Agents
There’s a quiet revolution happening in enterprise software development, and it’s not about the models. It’s about the harness. Personally, I think this is one of the most underreported shifts in AI adoption today. While everyone’s obsessing over the latest large language models (LLMs) from Anthropic, OpenAI, or Google, companies like Coinbase, Shopify, and Ramp are quietly building something far more strategic: their own coding agents. But here’s the twist—they’re not replacing commercial tools like Claude Code or Codex. Instead, they’re building the infrastructure around them.
What makes this particularly fascinating is the way enterprises are redefining the build-versus-buy debate. In my opinion, this isn’t about owning the model; it’s about owning the context. The agent harness—the execution environment that manages workflows, permissions, and integrations—is where the real value lies. It’s like building a custom dashboard for a high-performance car. The engine (the LLM) is powerful, but it’s the dashboard (the harness) that lets you control where you’re going.
The Harness: The New Strategic Layer
One thing that immediately stands out is how these companies are converging on a similar architecture. Coinbase’s Forge, Shopify’s River, and Ramp’s Inspect all solve remarkably similar problems, yet they’re built independently. What this really suggests is that there’s a repeatable pattern emerging. LangChain’s open-sourcing of Open SWE confirms it: enterprises are insourcing the harness while outsourcing the reasoning engine.
From my perspective, this is a natural evolution of how companies approach infrastructure. Think about cloud adoption a decade ago. Few enterprises built their own cloud from scratch. Instead, they built platforms on top of AWS or Azure to standardize deployments, security, and operations. AI is following the same playbook. The LLM becomes just another infrastructure dependency, while the harness becomes the control plane.
Why the Harness Matters More Than the Model
What many people don’t realize is that the competitive advantage in AI isn’t in the model itself—it’s in how you use it. Shopify’s River, for example, isn’t just an AI agent; it’s a platform that integrates with their monorepo, sandboxed environments, and credential management. This isn’t just about writing code; it’s about orchestrating workflows. River now participates in one out of every eight merged pull requests at Shopify. That’s not just efficiency—it’s a cultural shift.
If you take a step back and think about it, this is about control. By owning the harness, companies like Coinbase can optimize costs, enforce security policies, and experiment with models without disrupting developers. Chintan Turakhia from Coinbase mentions how they’ve reduced spending while processing more tokens. That’s only possible because they own the gateway between developers and models.
The Complementary Role of Commercial Tools
Here’s a detail that I find especially interesting: despite building their own agents, these companies still rely on commercial tools like Claude Code. Why? Because internal agents and commercial assistants serve different workflows. Internal agents handle asynchronous tasks—like converting bug reports into pull requests—while commercial tools dominate interactive sessions in editors or terminals.
This raises a deeper question: are we looking at a future where enterprises and vendors coexist in a symbiotic relationship? I think so. Model providers will compete to be the preferred reasoning engine inside platforms they don’t control. It’s a shift from selling models to selling compatibility.
The Unpredictable Economics of AI
AI economics remain a wild card. Walmart, Uber, and GitHub have all faced budget overruns due to unpredictable token consumption. A recent study from Stanford and Microsoft Research found that autonomous coding workflows can consume up to 1,000 times more tokens than interactive tasks. What this implies is that cost optimization isn’t just a model problem—it’s a platform problem.
This is where the harness becomes critical. By centralizing routing, caching, and model selection, platform teams can manage costs without burdening developers. It’s not just about saving money; it’s about making AI economically sustainable.
The Platform as the Strategic Asset
If there’s one takeaway, it’s this: the platform is the new strategic asset. A decade ago, companies differentiated themselves through deployment pipelines and infrastructure automation. Today, it’s about AI-assisted development. The LLM is becoming a commodity, while the harness—the platform—is where innovation happens.
In my opinion, this is the defining architectural shift of enterprise AI. The model providers will compete for a spot in platforms they don’t own. The real power lies in the layer that decides how AI is used, not the AI itself.
So, the next time someone asks you whether to build or buy an AI coding assistant, tell them it’s the wrong question. The real question is: Do you own the harness? Because that’s where the future of enterprise AI is being built.
