AI agents are reshaping how software is written, scaled, and experienced, and many expect the technology to unlock the gains AI firms have long promised. While most companies today remain in the “testing” phase, as agents make their way throughout the organization, workers will need to figure out how to integrate them into their workflows. That’s particularly true of developers, who can use agents to boost efficiency and in many cases will be also responsible for building, maintaining, and integrating them.
Agents are autonomous programs relying on underlying AI models like language models or planning systems that are capable of executing tasks without constant human orchestration. (As Chip Huyen has pointed out, many consider them “the ultimate goal of AI.”) It might sound obvious, but what distinguishes this as a novel approach is “agency”: operating independently according to preestablished goals, memory, and tools.
Agents can be simple, making a single API call based on user input, or complex, orchestrating multiple services, collaborating with other agents, and learning over time. But they’ll only ever be as useful as the data and systems they connect to, and that means that APIs will continue to play an outsize role. As the bridge between agents and the digital world, APIs make it possible for AI agents to access data, perform actions, and integrate with external systems to achieve their goals. But what does it mean to build for a world where agents, fueled by APIs, act on their own?
APIs aren’t a new technology; the concept dates back to the 1940s. And AI hasn’t changed the objective of a well-thought-out API: easily delivering valuable functionality to third parties. However, traditional APIs have always been designed with human developers in mind. Agent-compatible APIs don’t have the same requirements. For APIs to effectively serve agents, they need to be machine-consumable, self-describing, and semantically rich. This requires developers to prioritize clear functionality, descriptive metadata, and real-time error handling, all while maintaining accessibility for human users. There are also new protocols to consider, including the Model Context Protocol (MCP) and the Agent2Agent Protocol (A2A), which can be used to communicate with external data sources, tools, and other agents.
APIs aren’t going away any time soon, but developers intent on optimizing their systems and software should learn the new protocols that will help them connect agents with their systems and data. They also must consider the technical environment in which their APIs now circulate and design for both humans and agents. There’s no time like the present to get started.
Want to learn more? Join host Mike Amundsen and an esteemed lineup of API experts on July 17 for the O’Reilly API Superstream, all about creating APIs optimized for AI agents. Over four packed hours, you’ll explore issues with current APIs; how to integrate your APIs with AI and MCP; enterprise-grade agentic ecosystems; the synergy between APIs, LLMs, and XAI; Azure API Management; and much more. It’s free for O’Reilly members. Register here.
Not a member? Sign up for a free 10-day trial to attend—and check out all the other great resources on O’Reilly.
