Apps must be built for agents

Software teams are starting to build apps for two users at once: people on screens and agents calling tools in the background. (openai.com) That shift shows up in the interfaces themselves. OpenAI says its Responses API is “recommended for all new projects” and packages web search, file search, computer use, code interpreter, custom functions, and remote Model Context Protocol servers into one agent-oriented flow. (openai.com) The basic pattern is simple: the model does not just answer with prose, it asks to use a tool with named arguments, and the app executes that call. Google’s Gemini docs describe function calling as the bridge between natural-language requests and actions like scheduling meetings, creating invoices, or sending emails. (google.com) That changes what “good app design” means. Instead of a polished button for a human, developers now need a stable contract a model can follow every time, with fixed field names, required parameters, and outputs that parse cleanly. (openai.com) OpenAI’s structured outputs docs frame that as JSON Schema: a response format that forces required keys and valid enum values, so the app does not have to retry malformed answers. The same page says refusals can be detected programmatically, which lets developers route failures without guessing from free text. (openai.com) A second layer is the connection standard. Anthropic introduced the Model Context Protocol, or MCP, on November 25, 2024, as an open way for assistants to connect to content repositories, business tools, and development environments instead of relying on one-off integrations. (anthropic.com) MCP’s own architecture docs describe a host, a client, and a server: the app manages connections, the client maintains them, and the server exposes context or tools locally or over HTTP. In practice, that gives developers a common way to plug an agent into systems like files, databases, GitHub, Slack, or Stripe. (modelcontextprotocol.io; openai.com) The safety work moves down a level too. OpenAI’s help docs say its Audit Log API gives security teams an immutable record of events such as API key changes and account actions, and Gemini’s function-calling docs now assign each function call a unique ID so tool results can be matched back to the exact call. (openai.com; google.com) The documentation is changing with the software. Alongside API references, some teams now publish machine-readable guides like `llms.txt`, a proposed standard for telling language models which pages matter most and how to use a site at inference time. (llmstxt.org) The result is an app stack that looks less like a website and more like a controlled workbench for software that can read, decide, and act. The teams hiring for that stack are increasingly asking for the same things the platforms now document: tools, schemas, permissions, and logs. (openai.com; anthropic.com)