Integrated Labs versus Lightweight Orchestrators
Hark
The key advantage is economic flexibility, because orchestration startups can buy intelligence wholesale instead of funding the most expensive layer themselves. Companies like Browser Use and Browserbase package browser control, retries, and session handling around third party models, so when Claude, OpenAI, or another model improves, they can swap the engine without rebuilding the whole product. That keeps R&D and GPU spending far below a frontier lab that must train and refresh its own models.
-
In practice, these products sell the glue code around the model. They run browser sessions, keep login state alive, recover from broken page elements, and route tasks to a selected model. Browserbase even exposes model routing as part of Stagehand, which shows how value is shifting into the execution layer.
-
Anthropic carries a very different cost base. It is improving computer use inside Claude itself, and after acquiring Vercept in March 2026, it tied stronger execution capability directly to its own model stack. That can produce tighter reliability, but it also means bearing the fixed cost of frontier model training.
-
This is the same split seen across cloud software. One layer sells raw compute and models, the next layer sells workflow software on top. Hark is betting that owning both model behavior and execution harness creates a better agent, while open orchestration players are betting customers will prefer mixing and matching the best model each quarter.
Going forward, the market is likely to separate into expensive integrated labs and lighter agent software companies. If frontier models keep getting better quickly, model agnostic orchestrators gain leverage. If reliability becomes the deciding feature, companies that train the model and the action loop together, like Hark, have a clearer path to defend margin and product quality.