How Fieldway Built a Product Research Agent With Tavily

The problem

Matthew Stublefield, CEO of Fieldway and a longtime product leader, did not want an answer engine. He wanted to build one: an AI system that could run a full research plan, pull from the open web, and hand back synthesis he could put in front of a client.

The consumer AI tools he started with were built to answer you, not to be built on. He was stuck pasting prompts one at a time and babysitting the output. A cycle took about 5 hours, and that was the fast version. The slow version was 3 weeks of reading papers himself, or a six-figure research firm that took 1 to 2 months. Worse, the answers were drifting, with what he measured as a hallucination rate near 50%. "I don't want bite size," he said. "I want depth."

The solution

Matthew rebuilt his research on Tavily, the web layer purpose-built for AI agents. He wired it in through Tavily's MCP server in Claude Code in 1 to 2 days, turning his manual routine into an 8-step research agent. Now the agent runs the whole cycle: it prompts Tavily for live, cited web sources, pulls hundreds of grounded pages into his analysis, and verifies every citation with a Haiku model before he ever sees it. A cycle that once took 5 hours now runs in about 25 minutes. "Out of 400 to 500 citations, it's throwing out single digits," he said.

The wins followed fast. On one project, a fintech advisor brought Fieldway in to run the competitive analysis for a bank that was losing customers and could not pinpoint why. Matthew fed his agent 40-plus research documents, the bank's 16,000-row CRM, its sales collateral, and customer interview transcripts. Tavily-grounded desk research surfaced a whole class of competitor the bank had never tracked, the gap behind the churn. The work took 10 hours instead of the 3 to 4 weeks it once would have. The advisor's client was "thrilled" and "over the moon." Her reaction, in Matthew's words: "That's amazing. I want to hire you on a retainer to just keep doing that."

Why Tavily

Matthew ran the two side by side, and accuracy decided it. "I'm at a 50% hallucination rate with Perplexity," he said, while Tavily's sources held up under his Haiku verification. It was faster and deeper, and unlike a tool he queried by hand, its Search and research APIs wired straight into his agent through Claude Code. He has since moved all of his research onto Tavily, and every new client engagement runs on it. "It feels like it's being built for me," he said.

"I'm now getting something better than what those six-figure firms produced, in anywhere between 25 minutes and 2 hours." Matthew Stublefield, CEO, Fieldway

What's next

The pattern Matthew proved out, an agent that plans research, grounds it on Tavily, and verifies every citation, is what any team building production AI research reaches for. He is now extending it into a managed intelligence service, and his Tavily usage grows with every engagement he takes on. The systems are already built.

Matthew shared his setup and results for this story and is glad to compare notes with other teams weighing the same move.

See what Tavily could do for your team

If you want your agents grounding their outputs in real-time retrieval and cited sources instead of stale training data, reach out and we'll map it to your use case. Contact sales