Aparna Dhinakaran, writing on X in July 2026, argued that the software wrapping a coding model, not the model itself, is becoming the place where competitive advantage actually lives. Her term for that wrapper is the harness: the control loop that manages tool calls, routes tasks between models, sequences multi-step workflows, and decides what context the model sees at each turn. Her point is structural, not sentimental: coding models are converging on similar benchmarks fast enough that the model choice underneath an agent product matters less each quarter.
That framing forces a choice builders have mostly avoided naming directly. A harness built and shipped by the model’s own maker, tightly wired to that model’s context format, tool-calling conventions, and internal routing, delivers the best measured performance today. It knows its own model’s quirks natively and gets tuned in lockstep with every model update. The cost of that performance is coupling: the harness only fully works with one vendor’s model, and its advantages erode or reverse the day a competing lab ships something better.
The alternative is a portable, model-agnostic orchestration layer built independently of any single model maker. It talks to multiple models through a common interface, routes a task to whichever model is cheapest or best suited that week, and keeps the surrounding workflow logic, memory, tool definitions, and evaluation harness stable while the underlying model gets swapped out. Dhinakaran’s argument is that this portability, not raw performance, is the more durable asset as models keep changing and their prices keep falling.
Two concrete categories illustrate the split. Model-native agent runtimes, the coding-agent products shipped directly by the labs that train the underlying model, sit on one side. Model-agnostic orchestration layers, independent frameworks and internal tooling that treat the model as a swappable component behind a stable interface, sit on the other. Neither category is inherently correct. The tradeoff is peak performance now against resilience to churn later, and Dhinakaran frames it as a real bet, not a solved question.
The mechanism behind the durability claim is straightforward. If model prices keep dropping and capability keeps converging, the entity that owns the orchestration logic, not the model subscription, captures the compounding value: every workflow improvement, every tool integration, and every routing rule survives a model swap instead of resetting with it. A team locked into a vendor-native harness inherits that vendor’s roadmap, pricing, and deprecation schedule as its own.
The post did not include a specific timeline for when model convergence would make harness ownership decisive, and Dhinakaran did not attach numbers to the claim. The thesis is directional, not quantified, and builders should treat it that way rather than as a forecast with a date attached.
For the next ninety days, builders shipping agentic coding products should audit how much of their workflow logic is hardcoded to one vendor’s tool-calling format versus abstracted behind an interface that could route to a different model without a rewrite. Teams that cannot answer that question quickly have already made the vendor-lock bet, whether or not they meant to.
Based on commentary from Aparna Dhinakaran, writing on X in July 2026.
