The bottom line: Anthropic introduces Tool Search, Programmatic Tool Calling, and Tool Use Examples. These features enable AI agents to work with thousands of tools without exhausting context. Internal tests show significant improvements in memory efficiency and error reduction.
Anthropic unveils three new capabilities that enable AI agents to seamlessly work across hundreds or thousands of tools without overwhelming context memory. The innovations address core challenges in agent development and unlock entirely new use cases.
The future of AI agents lies in models operating effortlessly across hundreds or thousands of tools. An IDE assistant could integrate Git operations, file manipulation, package managers, test frameworks, and deployment pipelines. An operations coordinator could simultaneously connect Slack, GitHub, Google Drive, Jira, enterprise databases, and dozens of MCP servers.
Until now, a central challenge has been that tool definitions consumed enormous amounts of context tokens. A five-server setup can already consume 55,000 tokens: GitHub with 35 tools (26K tokens), Slack with 11 tools (21K tokens), Sentry with 5 tools (3K tokens), Grafana with 5 tools (3K tokens), and Splunk with 2 tools (2K tokens). Add to this frequent errors in incorrect tool selection and incorrect parameters, especially among similarly named tools.
Anthropic responds with three innovative features: The Tool Search Tool enables Claude to search for tools strategically, making thousands of tools accessible without overwhelming the context window. Programmatic Tool Calling allows Claude to invoke tools in a code execution environment, thereby reducing impact on the model’s context window. Tool Use Examples provides a universal standard for demonstrating effective tool usage.
Internal tests revealed impressive results from these features. Claude for Excel uses Programmatic Tool Calling to read and edit tables with thousands of rows without overwhelming the model. With the Tool Search Tool, 191,300 tokens of context are preserved, compared to 122,800 tokens with Claude’s traditional approach.
Source: www.anthropic.com