OpenAI functions agent in LangChain - Step-by-Step Execution

Concept Flow - OpenAI functions agent

User Input

↓

Agent Receives Input

↓

Agent Checks for Function Call

↓

Call Function

↓

Function Executes

↓

Return Function Output

↓

Agent Incorporates Output

↓

Respond to User

The agent receives user input, decides if a function call is needed, executes it if yes, then responds with the function output or text.

Execution Sample

LangChain

from langchain.agents import create_openai_functions_agent, AgentExecutor

agent = create_openai_functions_agent(llm, tools)
agent_executor = AgentExecutor(agent=agent, tools=tools)
response = agent_executor.invoke({'input': 'What is the weather in Paris?'})

This code creates an OpenAI functions agent with executor and sends a user question to get a response, possibly calling a function.

Execution Table

Step	Action	Input	Agent Decision	Function Called	Function Output	Agent Response
1	Receive user input	What is the weather in Paris?	Check if function needed
2	Decide to call function	What is the weather in Paris?	Yes, call weather function	get_weather(location='Paris')
3	Execute function	location='Paris'			Sunny, 25°C
4	Incorporate function output	Sunny, 25°C				The weather in Paris is Sunny, 25°C.
5	Return response to user					The weather in Paris is Sunny, 25°C.

💡 Agent returns final response after function call and output incorporation.

Variable Tracker

Variable	Start	After Step 2	After Step 3	After Step 4	Final
user_input		What is the weather in Paris?	What is the weather in Paris?	What is the weather in Paris?	What is the weather in Paris?
agent_decision		Check if function needed	Call function	Function output received	Response ready
function_called			get_weather(location='Paris')	get_weather(location='Paris')	get_weather(location='Paris')
function_output				Sunny, 25°C	Sunny, 25°C
agent_response					The weather in Paris is Sunny, 25°C.

Key Moments - 3 Insights

How does the agent know when to call a function instead of just replying?

The agent checks the input and matches it to available functions. If a function fits the request, it chooses to call it (see execution_table step 2).

What happens if the function output is empty or missing?

The agent waits for the function output before responding. If missing, it may fallback to a text response or error handling (not shown here).

Can the agent call multiple functions in one interaction?

Yes, but this example shows a single function call. The agent can chain calls by repeating the decision and execution steps.

Visual Quiz - 3 Questions

Look at the execution_table, what function does the agent call at step 2?

A send_email(address='Paris')

B get_news(topic='Paris')

C get_weather(location='Paris')

D No function called

💡 Hint

Check the 'Function Called' column at step 2 in the execution_table.

At which step does the agent receive the function output?

A Step 3

B Step 1

C Step 4

D Step 5

💡 Hint

Look at the 'Function Output' column in the execution_table.

If the user input was 'Tell me a joke', what would likely change in the execution_table?

A Agent calls a joke function instead of weather

B Agent skips function call and replies directly

C Agent calls get_weather function anyway

D Agent crashes

💡 Hint

Think about when the agent decides to call a function or just reply (see agent decision in execution_table step 2).

Concept Snapshot

OpenAI functions agent:
- Receives user input
- Checks if a function call fits the request
- Calls the function if needed
- Gets function output
- Responds using output or text
- Enables dynamic, real-time info in conversations

Full Transcript

The OpenAI functions agent works by first receiving what the user says. It then decides if it needs to call a special function to answer. If yes, it calls that function with the right details. The function runs and sends back an answer. The agent uses this answer to reply to the user. If no function is needed, the agent just replies with text. This way, the agent can give smart answers using live data or tools.