Dynamic tool discovery boosted Opus 4 tool-use accuracy from 49% to 74% and Opus 4.5 from 79.5% to 88.1%.
Engineering BlogThree new beta features for dynamic tool discovery: Tool Search Tool discovers tools on-demand instead of loading all upfront. Opus 4 improved from 49% to 74%, Opus 4.5 from 79.5% to 88.1% with Tool Search enabled.
Different calling patterns for tools depending on task structure. Sequential patterns call one tool, use its output for the next. Parallel patterns call multiple independent tools simultaneously. Recursive patterns call the same tool on intermediate outputs until convergence. Advanced Tool Use enables agents to choose appropriate patterns based on task structure rather than using all tools upfront.
Models generate structured tool search queries and structured tool call sequences that can be validated and optimized. Rather than unstructured natural language deciding which tools to use, Advanced Tool Use enforces structured specifications that reduce errors and enable validation. This is critical for reliable tool composition in agentic systems.
When a tool call fails or produces unexpected output, agentic systems need recovery mechanisms. Advanced Tool Use includes patterns for handling tool errors gracefully — retrying with different parameters, falling back to alternative tools, or reporting to the user. Robust error handling is the difference between tools that work 90% of the time and tools that work reliably in production.
Coordinating multiple tools across a workflow to achieve a complex goal. This requires the agent to understand tool dependencies (which tools must run before others), determine optimal calling order, handle partial failures, and aggregate results. Advanced Tool Use makes this tractable by providing dynamic discovery and structured selection patterns.