Talented individuals often rise to management positions, and this trend extends to AI models. Claude handles code execution, Gemini directs requests across CRM and chat systems, and GPT-5 coordinates public stock research. A significant shift in tool calling accuracy has occurred. Two years ago, GPT-4 managed less than 50% success in function-calling tasks, plagued by hallucinations and miscommunication. Today, state-of-the-art models achieve over 90% accuracy on these benchmarks, with models like Gemini 3 performing even better in real-world applications. The need for trillion-parameter models becomes clearer. Experiments with smaller models designed solely for tool selection often fail because they lack the necessary context. Management in AI requires an understanding of the world, which these smaller models can't provide. Currently, orchestrators can self-generate subagents for tasks, but this approach won't last. Larger models are better at handling complex requirements, but economic factors encourage smaller, efficient models. Techniques like distillation and reinforcement fine-tuning allow for models that are 40% smaller and 60% faster while retaining 97% of their performance. A new trend is emerging where specialized agents from various vendors are taking shape. The leading frontier model acts as an executive, directing requests to these specialists, which can include third-party vendors competing for excellence in their specific areas. As tool calling improves, teams can shift from monolithic models to specialized systems, boosting overall capabilities. The orchestration layer will likely be dominated by frontier labs, but they won't be able to control all the specialists. Startups focusing on niche applications, like advanced browser agents or business intelligence tools, can integrate into these specialized networks and carve out their own spaces in the market.