Building an Autonomous GPT-5 Agent

Introduction & Overview

Throughout this course, you have mastered the fundamentals of tool integration with GPT-5 : creating tool schemas, understanding GPT-5's tool use responses, and executing single tool requests. However, the approach you've learned so far has a significant limitation — it only handles one tool call per conversation turn. While this works perfectly for simple tasks, many real-world problems require multiple sequential steps, and often the number and nature of these steps cannot be determined in advance. In this lesson, we'll work together to transform GPT-5 from a single-turn tool user into an autonomous agent capable of iterative problem-solving. We'll build an agent class that can call tools, analyze results, decide what to do next, and continue this process until complex multi-step tasks are completed. This represents a fundamental shift from reactive tool usage to proactive, intelligent problem-solving that mirrors how humans approach complex challenges.

The Action-Feedback Loop Concept

Before we start coding, let's understand how autonomous agents operate through action-feedback loops, in which each tool execution provides information that influences the next decision. This iterative process mirrors human problem-solving: we take an action, observe the result, decide what to do next, and repeat until we reach our goal. The action-feedback loop consists of four key phases that repeat until task completion: Decision Phase: GPT-5 analyzes the current situation and determines the next action, which may include calling one or more tools. Action Phase: Our agent executes the requested tool(s) based on GPT-5's instructions. Feedback Phase: The results from the tool execution(s) are captured and added to the conversation history. Evaluation Phase: GPT-5 reviews the new information, decides whether the task is complete or if additional steps are needed, and the loop continues. This loop structure enables complex problem-solving because each iteration builds upon previous results. For example, when solving a quadratic equation, GPT-5 might first calculate the discriminant, then use that result to determine if real solutions exist, then calculate the square root of the discriminant, and finally compute the two solutions. The key insight is that GPT-5 doesn't need to plan all steps in advance — it can adapt its approach based on intermediate results, just like a human mathematician working through a problem. Now let's start building our agent class to make this iterative process possible.

Building Our Agent Class Foundation

Let's begin by creating the foundation of our autonomous agent. We need to establish the core structure that will manage extended conversations, tool execution, and decision-making loops. We'll start with the class definition and constructor: Pythonimport json from openai import OpenAI class Agent: # Base system prompt to be used for all agents BASE_SYSTEM_PROMPT = ( "You are an autonomous agent that can take multiple tool-calling steps when helpful. " "The user only sees your response when you stop using tools, not your tool usage or reasoning steps. " "When you provide your answer without calling tools, make it complete and standalone.\n" "Additional instructions:\n" ) def __init__( self, name, system_prompt="You are a helpful assistant.", model="gpt-5", tools=None, tool_schemas=None, max_turns=10, reasoning_effort="low" ): self.client = OpenAI() self.name = name self.model = model self.system_prompt = self.BASE_SYSTEM_PROMPT + system_prompt self.max_turns = max_turns self.reasoning_effort = reasoning_effort # Avoid shared mutable defaults and protect against external mutation self.tools = {} if tools is None else dict(tools) # name -> Python function self.tool_schemas = [] if tool_schemas is None else list(tool_schemas) # list of {type, name, description, parameters}import json from openai import OpenAI class Agent: # Base system prompt to be used for all agents BASE_SYSTEM_PROMPT = ( "You are an autonomous agent that can take multiple tool-calling steps when helpful. " "The user only sees your response when you stop using tools, not your tool usage or reasoning steps. " "When you provide your answer without calling tools, make it complete and standalone.\n" "Additional instructions:\n" ) def __init__( self, name, system_prompt="You are a helpful assistant.", model="gpt-5", tools=None, tool_schemas=None, max_turns=10, reasoning_effort="low" ): self.client = OpenAI() self.name = name self.model = model self.system_prompt = self.BASE_SYSTEM_PROMPT + system_prompt self.max_turns = max_turns self.reasoning_effort = reasoning_effort # Avoid shared mutable defaults and protect against external mutation self.tools = {} if tools is None else dict(tools) # name -> Python function self.tool_schemas = [] if tool_schemas is None else list(tool_schemas) # list of {type, name, description, parameters} Our agent's foundation relies on key design decisions that enable autonomous behavior while maintaining flexibility for different use cases: BASE_SYSTEM_PROMPT: Explicitly tells GPT-5 that it can make multiple tool calls and that users won't see the intermediate steps — only the final result. We're combining this with custom instructions to allow for domain-specific guidance while maintaining the autonomous behavior. Constructor parameters: Provide flexibility for different scenarios while ensuring some safe defaults: name: Provides a clear identifier for the agent, useful for debugging, logging, and when working with multiple agents in complex systems. system_prompt allows customization for specific domains like math or data analysis. model specifies which GPT-5 model to use for agent interactions. tools and tool_schemas: Default to None to avoid shared mutable defaults. Copied via dict(...) and list(...) so each agent gets its own independent registry and schema list, and to prevent later external mutations from affecting the agent. Accept any mapping/iterable-compatible inputs for convenience. max_turns prevents infinite loops by limiting the number of iterative steps. reasoning_effort controls how much computational effort GPT-5 applies to problem-solving, with options like "low," "medium," or "high." This architecture separates concerns cleanly while preparing us to implement the core functionality that will make our agent truly autonomous.

Implementing Tool Execution

Now let's add the method that handles individual tool executions within our agent loop. This method needs to be robust because tool failures shouldn't break our entire autonomous process: Pythondef call_tool(self, function_call): # Get the function name, arguments, and call ID tool_name = function_call.name call_id = function_call.call_id # Parse arguments from JSON string to Python dict tool_input = json.loads(function_call.arguments) # Display which tool is being called print(f"🔧 Tool called: {tool_name}({tool_input})") try: # Execute the tool with the given input result = self.tools[tool_name](**tool_input) except KeyError: # Return an error message if the tool is not found result = f"Error: Tool {tool_name} not found" except Exception as e: # Return an error message if the tool fails result = f"Error: {str(e)}" # Return the tool result in GPT-5 Responses API format return { "type": "function_call_output", "call_id": call_id, "output": json.dumps({"result": result}) }def call_tool(self, function_call): # Get the function name, arguments, and call ID tool_name = function_call.name call_id = function_call.call_id # Parse arguments from JSON string to Python dict tool_input = json.loads(function_call.arguments) # Display which tool is being called print(f"🔧 Tool called: {tool_name}({tool_input})") try: # Execute the tool with the given input result = self.tools[tool_name](**tool_input) except KeyError: # Return an error message if the tool is not found result = f"Error: Tool {tool_name} not found" except Exception as e: # Return an error message if the tool fails result = f"Error: {str(e)}" # Return the tool result in GPT-5 Responses API format return { "type": "function_call_output", "call_id": call_id, "output": json.dumps({"result": result}) } This method handles the individual tool executions that will happen within our larger iterative loop. Here's how it manages the execution flow: Extract tool information: Gets the tool name, call ID, and parses the arguments from JSON format into a Python dictionary that we can use to call the function. Debug tracking: Prints which tool is being called with what parameters — invaluable for debugging and understanding how our agent thinks. Execute with comprehensive error handling: Attempts to run the tool, catching both missing tools (KeyError — equivalent to checking if the tool exists in our dictionary) and execution failures (Exception). Return structured results: Converts both successful results and errors into properly formatted function call output objects with the correct call_id reference. The error handling ensures that tool failures don't break the entire agent loop. Instead, errors are converted into tool results that GPT-5 can understand and potentially work around. This robustness allows our agent to continue operating even when individual tools encounter problems, making the whole system much more resilient.

Building the Core Loop - Part 1: Understanding Stateless Design

Now we're ready to implement the heart of our autonomous agent: the run method. This method will manage the iterative loop that enables multi-step problem-solving. Let's start by understanding how our agent handles conversation state: Python def run(self, input_messages): # Create a copy of the input messages to avoid modifying the original messages = input_messages.copy() def run(self, input_messages): # Create a copy of the input messages to avoid modifying the original messages = input_messages.copy() The input_messages.copy() is important because it ensures our agent remains stateless . Just like normal LLM API calls, where you pass the complete conversation history each time, our agent doesn't store any conversation state between calls. Each time you call agent.run(), you provide the full context through input_messages, and the agent processes only that specific conversation without any memory of previous interactions. By copying the input messages instead of modifying them directly, we preserve the original conversation and allow the same agent instance to handle multiple independent conversations. This design also gives you complete control over context management — you can decide exactly what conversation history to include, filter out irrelevant messages, or combine conversations as needed before passing them to the agent.

Building the Core Loop - Part 2: Setting Up the Iteration

Now let's add the basic loop structure that will enable our agent's iterative problem-solving: Pythondef run(self, input_messages): # Create a copy of the input messages to avoid modifying the original messages = input_messages.copy() # Initialize turn counter to track iterations turn = 0 # Loop until the model returns a final answer or the max turns is reached while turn < self.max_turns: # Increment the turn turn += 1 # Ask the model for a response response = self.client.responses.create( model=self.model, instructions=self.system_prompt, input=messages, tools=self.tool_schemas, reasoning={"effort": self.reasoning_effort}, store=False )def run(self, input_messages): # Create a copy of the input messages to avoid modifying the original messages = input_messages.copy() # Initialize turn counter to track iterations turn = 0 # Loop until the model returns a final answer or the max turns is reached while turn < self.max_turns: # Increment the turn turn += 1 # Ask the model for a response response = self.client.responses.create( model=self.model, instructions=self.system_prompt, input=messages, tools=self.tool_schemas, reasoning={"effort": self.reasoning_effort}, store=False ) We're starting with a controlled loop that will continue until GPT-5 provides a final answer or we reach our maximum turn limit. Each iteration represents one complete action-feedback cycle in which GPT-5 makes a decision (potentially including tool calls), and we capture that decision in our conversation history. The turn counter prevents infinite loops while allowing sufficient iterations for complex problems. The API call uses the GPT-5 Responses API with several key parameters: instructions provides the combined base and custom instructions that guide the agent's behavior. input contains the full conversation history for context. tools specifies which tools are available for GPT-5 to use. reasoning controls the computational effort applied to problem-solving. store=False ensures we don't persist conversation data.

Building the Core Loop - Part 3: Handling Tool Calls

Now let's add the logic for handling tool calls within our loop. This is where the magic of autonomous behavior happens: Pythondef run(self, input_messages): # Create a copy of the input messages to avoid modifying the original... # Loop until the model returns a final answer or the max turns is reached while turn < self.max_turns: # Increment the turn... # Ask the model for a response... # Check if GPT wants to use any tools by looking for function_calls in output function_calls = [ item for item in response.output if item.type == "function_call" ] if function_calls: # Execute each function call function_outputs = [] # First, add the function calls to messages for function_call in function_calls: messages.append({ "type": "function_call", "name": function_call.name, "arguments": function_call.arguments, "call_id": function_call.call_id }) # Then execute and collect the outputs for function_call in function_calls: # Execute the tool with the given input tool_result = self.call_tool(function_call) # Add result to function outputs list function_outputs.append(tool_result) # Add all function outputs to messages messages.extend(function_outputs)def run(self, input_messages): # Create a copy of the input messages to avoid modifying the original... # Loop until the model returns a final answer or the max turns is reached while turn < self.max_turns: # Increment the turn... # Ask the model for a response... # Check if GPT wants to use any tools by looking for function_calls in output function_calls = [ item for item in response.output if item.type == "function_call" ] if function_calls: # Execute each function call function_outputs = [] # First, add the function calls to messages for function_call in function_calls: messages.append({ "type": "function_call", "name": function_call.name, "arguments": function_call.arguments, "call_id": function_call.call_id }) # Then execute and collect the outputs for function_call in function_calls: # Execute the tool with the given input tool_result = self.call_tool(function_call) # Add result to function outputs list function_outputs.append(tool_result) # Add all function outputs to messages messages.extend(function_outputs) When GPT-5 decides to use tools, we handle the execution through a systematic two-step process: Detect function calls: We filter the response output array to find all items with type == "function_call", which tells us GPT-5 wants to execute tools. Add function calls to conversation: Before executing anything, we add each function call to the messages array. This maintains a complete record of what GPT-5 requested. Execute all requested tools: GPT-5 might call multiple tools in a single turn, and we need to execute each one to gather all the information it needs for its next decision. Collect and add results: We execute each tool while collecting the results, then add all function outputs to the messages array. This two-step approach (calls first, then outputs) maintains the proper conversation structure that GPT-5 expects. Each tool result influences GPT-5's subsequent reasoning, allowing it to build upon what it just learned and make more informed decisions in the next turn.

Building the Core Loop - Part 4: Managing Flow Control

Finally, let's complete our loop with the logic for handling final responses and error conditions: Pythondef run(self, input_messages): # Create a copy of the input messages to avoid modifying the original... # Loop until the model returns a final answer or the max turns is reached while turn < self.max_turns: # Increment the turn... # Ask the model for a response... # Check if GPT wants to use any tools by looking for function_calls in output function_calls = [ item for item in response.output if item.type == "function_call" ] if function_calls: # Execute each function call and add results to messages... else: # Add the final response to messages messages.append({ "role": "assistant", "content": response.output_text }) # Return the agent history and final output return messages, response.output_text # If the max turns is reached, raise an exception raise Exception("Max turns reached")def run(self, input_messages): # Create a copy of the input messages to avoid modifying the original... # Loop until the model returns a final answer or the max turns is reached while turn < self.max_turns: # Increment the turn... # Ask the model for a response... # Check if GPT wants to use any tools by looking for function_calls in output function_calls = [ item for item in response.output if item.type == "function_call" ] if function_calls: # Execute each function call and add results to messages... else: # Add the final response to messages messages.append({ "role": "assistant", "content": response.output_text }) # Return the agent history and final output return messages, response.output_text # If the max turns is reached, raise an exception raise Exception("Max turns reached") When GPT-5 reaches a final answer, we handle the completion through a structured return process: Detect completion: When GPT-5 doesn't want to use tools (no function calls in the output), it signals that it has reached a final answer and no further iterations are needed. Extract clean response: GPT-5 provides the final text directly through response.output_text, which contains the readable answer without any tool-related content. Add to conversation history: Before returning, we add the final assistant response to the messages array to maintain a complete conversation record. Return complete state: We return both the full conversation history (messages) and the final response text (response.output_text) to maintain our stateless design — the caller receives everything needed to understand what happened and can use the conversation history for follow-up questions or multi-turn interactions. Safety net for runaway loops: The exception for reaching max turns prevents infinite loops if something goes wrong. You can control this limit through the max_turns parameter, or alternatively implement a mechanism to force GPT-5 to provide a final answer when approaching the limit rather than raising an exception. This dual return approach reinforces our stateless architecture by giving the caller complete control over conversation state while providing both the detailed interaction history for continued conversations and the clean final answer for immediate use.

Complete Run Method

Here's how our complete run method looks when put together: Pythondef run(self, input_messages): # Create a copy of the input messages to avoid modifying the original messages = input_messages.copy() # Initialize turn counter to track iterations turn = 0 # Loop until the model returns a final answer or the max turns is reached while turn < self.max_turns: # Increment the turn turn += 1 # Ask the model for a response response = self.client.responses.create( model=self.model, instructions=self.system_prompt, input=messages, tools=self.tool_schemas, reasoning={"effort": self.reasoning_effort}, store=False ) # Check if GPT wants to use any tools by looking for function_calls in output function_calls = [ item for item in response.output if item.type == "function_call" ] if function_calls: # Execute each function call function_outputs = [] # First, add the function calls to messages for function_call in function_calls: messages.append({ "type": "function_call", "name": function_call.name, "arguments": function_call.arguments, "call_id": function_call.call_id }) # Then execute and collect the outputs for function_call in function_calls: # Execute the tool with the given input tool_result = self.call_tool(function_call) # Add result to function outputs list function_outputs.append(tool_result) # Add all function outputs to messages messages.extend(function_outputs) else: # Add the final response to messages messages.append({ "role": "assistant", "content": response.output_text }) # Return the agent history and final output return messages, response.output_text # If the max turns is reached, raise an exception raise Exception("Max turns reached")def run(self, input_messages): # Create a copy of the input messages to avoid modifying the original messages = input_messages.copy() # Initialize turn counter to track iterations turn = 0 # Loop until the model returns a final answer or the max turns is reached while turn < self.max_turns: # Increment the turn turn += 1 # Ask the model for a response response = self.client.responses.create( model=self.model, instructions=self.system_prompt, input=messages, tools=self.tool_schemas, reasoning={"effort": self.reasoning_effort}, store=False ) # Check if GPT wants to use any tools by looking for function_calls in output function_calls = [ item for item in response.output if item.type == "function_call" ] if function_calls: # Execute each function call function_outputs = [] # First, add the function calls to messages for function_call in function_calls: messages.append({ "type": "function_call", "name": function_call.name, "arguments": function_call.arguments, "call_id": function_call.call_id }) # Then execute and collect the outputs for function_call in function_calls: # Execute the tool with the given input tool_result = self.call_tool(function_call) # Add result to function outputs list function_outputs.append(tool_result) # Add all function outputs to messages messages.extend(function_outputs) else: # Add the final response to messages messages.append({ "role": "assistant", "content": response.output_text }) # Return the agent history and final output return messages, response.output_text # If the max turns is reached, raise an exception raise Exception("Max turns reached")

Testing Our Autonomous Agent

Now let's put our agent to work! We'll create a math-focused autonomous agent and see how it handles a complex quadratic equation. We'll provide more math tools following the same pattern used across the course, so you can easily extend your agent's capabilities as needed: Pythonimport json from agent import Agent from functions import sum_numbers, multiply_numbers, subtract_numbers, divide_numbers, power, square_root # Load the schemas from JSON file with open('schemas.json', 'r') as f: tool_schemas = json.load(f) # Create a dictionary mapping tool names to functions tools = { "sum_numbers": sum_numbers, "multiply_numbers": multiply_numbers, "subtract_numbers": subtract_numbers, "divide_numbers": divide_numbers, "power": power, "square_root": square_root } # Create a stateless autonomous agent agent = Agent( name="math_assistant", system_prompt="You are a helpful math assistant. Always use the available tools to perform calculations accurately.", tools=tools, tool_schemas=tool_schemas, max_turns=15 ) # Initialize conversation with user message messages = [{"role": "user", "content": "Solve this equation: 2x² - 7x + 3 = 0 using tools"}] # Send message to the stateless agent messages, result = agent.run(messages) # Display the response print("\nFinal response:") print(result)import json from agent import Agent from functions import sum_numbers, multiply_numbers, subtract_numbers, divide_numbers, power, square_root # Load the schemas from JSON file with open('schemas.json', 'r') as f: tool_schemas = json.load(f) # Create a dictionary mapping tool names to functions tools = { "sum_numbers": sum_numbers, "multiply_numbers": multiply_numbers, "subtract_numbers": subtract_numbers, "divide_numbers": divide_numbers, "power": power, "square_root": square_root } # Create a stateless autonomous agent agent = Agent( name="math_assistant", system_prompt="You are a helpful math assistant. Always use the available tools to perform calculations accurately.", tools=tools, tool_schemas=tool_schemas, max_turns=15 ) # Initialize conversation with user message messages = [{"role": "user", "content": "Solve this equation: 2x² - 7x + 3 = 0 using tools"}] # Send message to the stateless agent messages, result = agent.run(messages) # Display the response print("\nFinal response:") print(result) When we run this code, our agent demonstrates sophisticated autonomous reasoning: text🔧 Tool called: power({'base': -7, 'exponent': 2}) 🔧 Tool called: multiply_numbers({'a': 2, 'b': 3}) 🔧 Tool called: multiply_numbers({'a': 4, 'b': 6}) 🔧 Tool called: subtract_numbers({'a': 49, 'b': 24}) 🔧 Tool called: square_root({'number': 25}) Final response: Using the quadratic formula x = [-b ± sqrt(b^2 - 4ac)]/(2a) with a=2, b=-7, c=3: - Discriminant: b^2 - 4ac = 49 - 24 = 25 - sqrt(25) = 5 Thus: - x = (7 + 5)/4 = 12/4 = 3 - x = (7 - 5)/4 = 2/4 = 1/2 Solutions: x = 3 and x = 1/2.🔧 Tool called: power({'base': -7, 'exponent': 2}) 🔧 Tool called: multiply_numbers({'a': 2, 'b': 3}) 🔧 Tool called: multiply_numbers({'a': 4, 'b': 6}) 🔧 Tool called: subtract_numbers({'a': 49, 'b': 24}) 🔧 Tool called: square_root({'number': 25}) Final response: Using the quadratic formula x = [-b ± sqrt(b^2 - 4ac)]/(2a) with a=2, b=-7, c=3: - Discriminant: b^2 - 4ac = 49 - 24 = 25 - sqrt(25) = 5 Thus: - x = (7 + 5)/4 = 12/4 = 3 - x = (7 - 5)/4 = 2/4 = 1/2 Solutions: x = 3 and x = 1/2. Our agent systematically applied the quadratic formula by calculating b² ((-7)²), computing 4ac (first 2×3, then 4×6), finding the discriminant (49-24), and taking the square root (√25). Each tool call built upon previous results, demonstrating true autonomous reasoning. The agent made 5 tool calls across multiple conversation turns, yet the user only sees the final, complete answer. Note that GPT-5 may use different tool call sequences depending on the reasoning_effort parameter — higher effort levels may result in more thorough step-by-step calculations, while lower effort levels might optimize for efficiency. You can experiment with this parameter to find the right balance for your use case.

Summary & Practice Preparation

Together, we've successfully built an autonomous agent capable of complex, multi-step problem-solving. Our agent class encapsulates conversation management, tool execution, and iterative decision-making in a reusable structure that can tackle problems requiring dozens of sequential operations. The architecture we created enables GPT-5 to operate as a true autonomous agent: it can assess situations, make decisions, execute tools, learn from results, and continue iterating until complex tasks are completed. This represents a fundamental advancement from simple tool usage to intelligent, adaptive problem-solving. In the upcoming practice exercises, you'll implement your own autonomous agents, experiment with different instructions and tool combinations, and tackle increasingly complex multi-step problems. You'll gain hands-on experience with the debugging and optimization techniques needed for production agent systems, building upon the solid foundation we've created together.