PydanticAI - Building a Research Agent

A research agent built with PydanticAI can turn a single question into multiple targeted web searches, run them concurrently, and return a structured, markdown-ready report—while letting developers control output shape and inject time-sensitive context. The core payoff is reliability: instead of getting free-form text, the agent produces a predictable schema (title, main section, and bullet summaries) that can be rendered or further processed downstream.

The setup starts with choosing a search backend. DuckDuckGo is presented as a straightforward option but prone to rate limiting, especially without a proxy. Tavily is offered as a practical alternative, including a paid API tier that still provides “a thousand free calls a month,” making it easier to iterate without throttling. Both search paths support synchronous and asynchronous usage, but the agent is configured for asynchronous calls so multiple searches can run in parallel.

The agent’s structure hinges on three Pydantic-style building blocks. First, a small “search data” class carries parameters like max results per search. Second, a “result type” defines the final structured output: a research title, a research main section, and research bullets, all as strings intended for markdown formatting. Third, a system prompt instructs the model to generate strong keywords for 3–5 searches total, label each query with a query number, and then synthesize the retrieved information into the required schema. The prompt is intentionally adjustable; it can steer the agent toward academic or commercial framing depending on the use case.

A dedicated search tool is then created using a decorator-based interface. The tool accepts the model-generated query and query number, calls Tavily’s search context method, and returns results constrained by the max-results dependency. This tool wiring is what lets the model decide how many searches to run and what each query should be, while the developer keeps tight control over what the tool returns.

When asked for “a very detailed bio on Sam Altman,” the agent chooses four searches and runs them asynchronously—each search retrieving three results—before formatting everything into the structured markdown response. The transcript also highlights cost awareness: token usage can climb quickly because multiple searches generate a lot of text for the model to process. The agent can be tuned to be more economical by limiting searches or adjusting prompt guidance.

The most consequential refinement comes with time sensitivity. Asking for “the latest new AI news” initially yields outdated information because the system prompt lacks a “today’s date” reference; the model falls back to its cutoff (shown as October 2023). Injecting a “today” dependency into the system prompt fixes this: the agent now targets “the last few days,” decides it only needs three searches, and returns bullets aligned with early December 2024 items such as OpenAI’s “12 days of Shipmas” events and rumors about a ChatGPT Pro plan and the text2video SORA model.

Overall, the approach demonstrates the fundamentals needed for a research agent: structured output via a Pydantic schema, tool-based web retrieval, asynchronous multi-query execution, and dependency-driven prompt injection to keep results current. The next step foreshadows a similar agent pattern combined with RAG and a vector store for grounded answers from curated documents.