Building Custom Tools and Agents with LangChain (gpt-3.5-turbo)

Custom tools are the key lever for making LangChain conversational agents more useful—and the biggest practical lesson is that tool use often requires careful prompt and tool-description tuning, especially with chat-based models like gpt-3.5-turbo. The walkthrough starts with the simplest possible tools: a “meaning of life” function that returns a fixed string (42.17658) and a “random number” function that returns a random value. Both are wrapped as LangChain tools so a ReAct-style agent can decide when to call them, based on the tool name, description, and expected input.

When the agent is asked “What is the time in London?”, it correctly selects an external search tool (DuckDuckGo in the earlier context) and returns search results. Asking for “a random number” also works as expected: the agent chooses the random-number tool, receives the output (e.g., 4), and then responds in a chatty style that includes the tool result in its message history. But the “meaning of life” tool reveals a core failure mode: even when a tool exists and the question matches it, the chat model may still answer directly instead of calling the tool. In this run, the agent produced a philosophical explanation rather than returning the tool’s exact value.

The fix is not changing the tool logic—it’s changing the instructions. The transcript shows prompt engineering as a requirement: the system prompt is overwritten to explicitly tell the assistant it “does[es]n’t know” about meaning-of-life answers and must use the meaning-of-life tool for those topics. After that adjustment, the agent reliably calls the tool, uses the action input (“MOL”), and returns the tool’s full string. That also surfaces another nuance: the tool output includes more than the user asked for (it returns both the rounded and unrounded value), and the model may not automatically trim it to match the user’s wording. The takeaway is that tool descriptions and prompts must be precise about both when to call tools and what level of formatting the tool should return.

Next comes a more practical custom tool: a webpage “page getter” that fetches HTML, strips tags using BeautifulSoup, and truncates output to avoid context limits (the transcript notes a ~4,000+ token input constraint). The tool also acknowledges messy real-world text extraction—like excessive newlines—and suggests filtering them. Implemented either as a simple function tool or as a LangChain BaseTool subclass with a run method, it returns cleaned text to the agent.

With the updated prompt—now also instructing the assistant to use the webpage tool when asked about web content—the agent can answer questions like “Is there an article about Clubhouse on TechCrunch today?” and “What are the titles of the top stories on CBSnews.com.” In both cases, the agent fetches the relevant page (sometimes via an intermediate search step), extracts enough text to identify the requested items, and returns titles and dates. The overall message is clear: custom tools make agents dynamic, but reliable tool use depends on aligning prompts, tool descriptions, and output formatting with the model’s chat behavior and context constraints.