Nimble Web Search Agents targets AI research token costs

Nimble launched Web Search Agents, a new product for AI teams that need agents to run web research without sending broad, generic search results through a model. The company says the product learns a customer’s domain, selects retrieval strategies for that task and reduces the token waste that comes from processing irrelevant pages.

The release is aimed at enterprises and AI-native companies building production agents for work such as market research, lead enrichment, competitive monitoring and structured dataset generation. Nimble did not disclose pricing, revenue, valuation or headcount for the launch.

What is Nimble Web Search Agents?

Web Search Agents is a web retrieval layer for AI agents that can be accessed through Nimble’s API, software development kit and Model Context Protocol. Nimble says developers can add it as a tool inside an existing agent or use it to build applications ranging from lower-latency search to longer-running research workflows.

The company’s argument is that a general search tool is a poor fit for agents doing specialized business research. In Nimble’s view, a lead enrichment agent and a market research agent should not receive the same undifferentiated search output, because each task has different signals, sources and tolerance for noise.

Nimble says its system combines proprietary indexes with live retrieval from websites and checks its own work during the research process. That is a cost claim as much as a quality claim: fewer irrelevant sources should mean fewer tool calls and fewer tokens spent reading pages that do not contribute to the answer.

What did Nimble claim in benchmarks?

Nimble said its own benchmark testing found a 21-point improvement in answer quality and a 51% reduction in tokens used per query. The company did not provide a full public methodology in the announcement, so buyers should treat those numbers as vendor benchmarks rather than neutral market data.

Uri Knorovich, Nimble’s chief executive and co-founder, said enterprises are constrained by accuracy and cost when agents need current web context. He said the goal is to give agents the right live information without relying on generic search output or spending tokens unnecessarily.

Speed is not the central selling point. Nimble is positioning the product for research jobs that can run for hours and where missing a relevant source may matter more than returning a fast answer. That puts the product closer to enterprise research infrastructure than a consumer-style search interface.

Which customers are using it?

Rox, an AI-native customer relationship management company, said it cut token costs by 20-fold after using the service and improved the quality and completeness of the information reaching its agents. Qodo, a code integrity startup, said Nimble helped tune a Claude Managed Agent to find competitor signals relevant to its team rather than return a broad market overview.

Nimble says its platform handles more than 90 million searches per day and serves both Fortune 500 companies and younger AI-native businesses. The company was founded in 2021 and is based in New York.

The launch follows Nimble’s $47 million Series B round in February, led by Norwest Venture Partners. Databricks Ventures, Target Global Management, Square Peg Capital, Hetz Ventures, Slow Ventures, R-Squared Ventures, J-Ventures and InvestInData also participated. Nimble says it has raised $75 million to date.

This story draws on original reporting from SiliconANGLE.