The Next Hop for APIs: From Capability Interfaces to Scenario Interfaces

On April 8, 2026, Anthropic released Claude Managed Agents, a composable set of cloud agent APIs that provide sandboxed execution, persistent state, permission control, and automatic scaling. Developers no longer have to build agent infrastructure themselves and can create, run, and manage complete agent workflows through the API. Notion, Rakuten, and Asana are the first users. Anthropic says it can cut agent development from months to weeks. This is not an isolated product move. Over the past year and a half Anthropic has shipped MCP (a protocol for connecting tools), Computer Use (desktop operation), the Agent SDK (an agent development framework), Tool Search and Programmatic Tool Calling (tool management at scale), and now agent hosting and orchestration inside the API. Strung together, they point in one direction: the model API is changing from a capability interface into a scenario interface. What it returns is no longer a piece of text but a whole runnable scenario. APIs move up a level and the middle layer gets eaten Anthropic isn't alone. OpenAI's Responses API, launched in March 2025, pulled into the API layer everything developers had to build themselves in the Chat Completions era: server-side state, web search, code execution, and remote MCP servers. Calling GPT-5 through Responses scores 3 percentage points higher on SWE-bench than through Chat Completions, and cache utilization improves 40% to 80%. The Assistants API will be sunset in the first half of 2026, and Responses is the only recommended entry. Google goes further. The Interactions API it launched this year is more than an upgrade to generateContent. It is a unified interface for interacting with models and agents, with built-in state management, tool orchestration, and long-running task handling, and it can call compound agents such as Gemini Deep Research directly. The Gemini 3 Pro API already supports Computer Use, Thought Signatures, and adjustable reasoning depth natively. Three vendors, three paths, one end state: the API itself becomes an agent runtime. To build an agent that could search, run code, and remember context, developers used to stack LangChain, AutoGPT, or an in-house orchestration framework on top of a model API. Model vendors are now absorbing those capabilities into the API layer. The middle layer's core value was bridging the gap between model capability and application scenario, and when the gap is filled it gets squeezed structurally. MCP is very likely a transitional form