ML competition shifts to agents
Discussion in the feed says the competitive edge in ML is moving away from raw model size to how vendors lock access with agents and workflow integrations. (x.com) That framing is echoed by posts about architectures and automated agents being the next battleground for product moats, not just better base models. (x.com)
Machine learning vendors are increasingly competing on the software wrapped around models — agents, tools and workflow hooks — rather than on model size alone. (openai.com) An agent is a model that can take steps on its own: search the web, open files, call software tools, or click through a computer screen to finish a task. OpenAI’s Responses API documents built-in tools including web search, file search and computer use, and its Agents Software Development Kit says agents “plan, call tools, collaborate across specialists, and keep enough state” for multi-step work. (developers.openai.com 1) (developers.openai.com 2) That product layer is now explicit across major vendors. Google Cloud says Vertex AI Agent Builder is for building, scaling and governing “enterprise-grade agents” grounded in company data, while Microsoft says Copilot Studio agents can be extended with connectors to external services and application programming interfaces. (cloud.google.com) (learn.microsoft.com) Anthropic has made the same bet from a different angle. Its Claude computer-use system can move a cursor, click and type through software interfaces, and the company’s Model Context Protocol is an open standard for connecting assistants to business tools and data repositories. (anthropic.com 1) (anthropic.com 2) The practical effect is that the model becomes one layer in a larger stack. The harder part to copy is often the surrounding system: which internal data the agent can reach, which approvals it needs, which applications it can operate, and how deeply it is embedded in a company’s daily workflow. (docs.cloud.google.com) (learn.microsoft.com) That shift has been reinforced by the way vendors describe their own products in 2025 and 2026. OpenAI says developers can use the Agents Software Development Kit when their application owns orchestration, tool execution, approvals and state, while Google’s Agent Engine documentation focuses on developing and deploying agents into production systems. (developers.openai.com) (docs.cloud.google.com) The business logic is straightforward: a base model can be swapped more easily than a workflow tied into email, documents, customer records and internal permissions. Microsoft’s documentation says connectors can ground Copilot Studio agents in Microsoft 365, Dynamics 365, Microsoft Fabric and non-Microsoft enterprise data, which raises switching costs once those links are in place. (learn.microsoft.com 1) (learn.microsoft.com 2) There is still a counterargument from model makers: better base models improve every agent built on top of them. OpenAI’s migration guide says reasoning models in the Responses API deliver “better model intelligence,” and Anthropic’s engineering team says infrastructure setup can move agentic coding benchmark scores by several percentage points, suggesting both model quality and system design still matter. (developers.openai.com) (anthropic.com) The result is a competition that looks less like a race for the single biggest model and more like a race to own the task. In that market, the winning product is not just the system that answers a question, but the one that can finish the work inside the software people already use. (openai.com) (docs.cloud.google.com)
