The future of AI agents lies in models that integrate and operate effortlessly with hundreds or thousands of different tools. An IDE companion that seamlessly incorporates git workflows, file operations, package management, testing tools, and CI/CD pipelines. An operations coordinator that simultaneously integrates Slack, GitHub, Google Drive, Jira, company databases, and dozens of MCP servers. To create truly effective agents, they must be able to work with unlimited tool libraries without cramming every definition into the context from the start. Our blog post on MCP code execution highlighted how tool outputs and definitions can sometimes exceed 50,000 tokens even before the agent processes the actual request. Agents ought to dynamically discover and load only the tools needed for their current task. They should also be able to invoke tools directly from within code. With natural language tool calling, every invocation demands a complete inference pass, causing intermediate results to accumulate in the context regardless of their usefulness. Code works well for orchestration tasks like loops, conditional statements, and data transformations. Agents require the ability to flexibly switch between code execution and inference depending on the specific task. They must also learn proper tool usage from concrete examples, rather than relying solely on schema definitions.
Anthropic Engineering