OpenAI Functions + LangChain : Building a Multi Tool Agent

OpenAI’s function-calling system, wired through LangChain, can turn a plain chat model into a finance assistant that reliably selects the right API tool, extracts the right parameters, and returns a grounded answer. The core workflow is: define one or more callable functions (with names, descriptions, and a strict parameter schema), let the model decide when to invoke them, execute the tool with the model-provided arguments, then feed the tool result back so the model can produce the final natural-language response.

The example starts with a simple “finance bot” built on the Yahoo Finance API. Users ask questions like “What is the price of Google stock?” or “Has Apple gone up over the past 90 days?” The key challenge—users won’t know ticker symbols—is handled by the model’s built-in knowledge: when asked about Apple or Google, it outputs the correct Yahoo Finance ticker behind the scenes. In the manual setup, the assistant first sends a human message plus a list of function definitions to the model. The model responds not with an answer, but with a structured function call: the function name (e.g., “get_stock_ticker_price”) and JSON arguments (e.g., the ticker for Google). The code then runs the corresponding tool using those arguments, retrieves the real price, and returns that result to the model using a dedicated function message. Only after the tool output is provided does the model generate the final response such as “The current price of Google stock is 123.83.”

LangChain then streamlines this “long way” by using an agent type designed for OpenAI functions. Instead of manually converting tools into OpenAI function formats and manually routing messages, the agent handles tool selection, argument passing, tool execution, and response synthesis. This approach is presented as an advantage over older prompt-based patterns (like ReAct-style tool prompting): it tends to improve tool selection and reasoning while reducing token waste from heavy in-context examples. Tradeoffs remain: customization is less straightforward than prompt tinkering, the setup is currently more tightly coupled to OpenAI’s function-calling conventions, and tool descriptions/schemas still consume tokens.

The finance bot expands from one tool to multiple tools. One function computes percentage price change over a time window given a ticker and a number of days; another finds the best-performing stock among a list of tickers over a specified period. The model can interpret user time expressions and convert them into the tool’s expected inputs—asking for “three months” maps to “90 days,” while “a month” maps to a shorter day count. It also handles mixed ticker formats: users can provide full company names (Google, Meta, Microsoft) and the model supplies the correct Yahoo Finance tickers. The same mechanism even works for crypto comparisons; when asked about Bitcoin over three months, the model uses Yahoo’s expected symbol format (not just a generic “BTC”), enabling comparisons across stocks and cryptocurrencies.

Overall, the transcript demonstrates a practical recipe for building multi-tool agents: strict schemas for tool inputs, function-call orchestration, and agent-based automation in LangChain—resulting in a conversational system that can answer finance questions grounded in external APIs.