How they use Tavily

IBM's Software Innovations Lab uses Tavily for web search in CUGA, its open-source generalist agent. Tavily is included as a default tool, but search in CUGA's core agent is optional and starts switched off. The planner calls Tavily as a tool and can request another search before answering. An IBM engineer identified Tavily as the team's search tool for customizing CUGA agents. The team plans to expand its Tavily use as agent adoption grows.

The team also built a shared Tavily-backed MCP server on IBM Code Engine for two dozen applications in the CUGA gallery. This keeps the API key server-side. Tavily replaced page scraping and supplies readable content, relevance scores and responses that fit the agent's tool schema without an adapter. Relevance scores guide result truncation to fit the context budget. Web access is controlled through an app-definition flag, allowing policy approval without code changes. The gallery's Web Researcher app prevents fallback to the model's training data. It runs two to four searches per question, produces a report with a confidence rating and cites a source for every claim. Ouroboros, a seven-agent lead-generation app, uses Tavily search in its person finder and passes result content to a step that infers likely email patterns.

IBM and Tavily have also published an open-source spreadsheet-enrichment integration that combines Granite models with Tavily search APIs as the first step in a strategic partnership. Separately, IBM selected Tavily's Research Agent for a limited watsonx Orchestrate eCommerce pilot and listed it in its agent catalogue. IBM MediaCenter also presents a watsonx Orchestrate partner-agent demo featuring Tavily Research Agent. These are distinct IBM projects.