LangChain Agents - Joining Tools and Chains with Decisions

LangChain agents let a language model choose—at runtime—which tools to use (or whether to use any tools at all) to answer a user’s question. Instead of forcing every query through a single tool or a fixed chain, an agent runs an “executor” step that reads the input, decides the best next action based on the tools it was initialized with, and then iterates through tool calls until it can produce a final response. That decision-and-action loop is what turns a collection of tools (search, calculator, Wikipedia, terminal) into a flexible assistant that can handle mixed questions.

The walkthrough starts with an agent configured with two tools: a search tool backed by a search API and a calculator tool for math. The agent is initialized with a language model (OpenAI, temperature set to 0) and an agent type called “zero-shot react.” Initialization also constructs the prompt the model uses to plan: it’s instructed that it has access to specific tools, must follow a structured format, and should generate intermediate “thought” steps that determine which action to take and what input to send to that action. When prompted with a casual question like “How are you today?”, the agent produces a direct response without calling any tool. When asked a question requiring external facts and computation—such as “Who is the United States president? What is his current age divided by 2?”—the agent uses search to identify the president (Joe Biden), uses search again to obtain Biden’s age (80), then calls the calculator to compute 80/2 (40), and finally returns the composed result.

The agent’s ability to string multiple tools together becomes clearer with a compound query: finding the average age in the United States and comparing it to the current U.S. president’s age. It searches for the needed statistics, searches again for the president’s age, then synthesizes the final comparison.

Adding more tools expands what the agent can do. A second agent includes Wikipedia and a terminal tool in addition to search and calculator. For example, it looks up “head of DeepMind” via search, then uses Wikipedia to define “DeepMind” and identify it as a British AI research lab founded in 2010 and a subsidiary of Alphabet. When asked to add 50 years to DeepMind’s founding year and predict whether AGI will exist by then, it retrieves the founding year from Wikipedia and then performs the arithmetic without invoking the calculator tool—suggesting that some simple computations can be handled directly by the language model. When asked for DeepMind’s office location, it again uses search, and then can perform math on the retrieved address (squaring the street number) by calling the calculator.

The terminal tool demonstrates another capability: inspecting the local environment. Asked what files exist in the current directory, the agent runs an `ls` command, navigates to a sample data folder, and lists available files. It can also search within files using commands like `grep` to answer questions such as whether a file about “California” exists. The walkthrough ends with a caution: terminal access can be dangerous if it allows destructive commands, so it should be added carefully and ideally constrained to safe operations.