Minimal AI agent tutorial

Skip to content

Initializing search

swe-agent/mini-swe-agent

Let's make it more robust

mini-swe-agent

Contribute to this guide

Comments

Let's make it more robust

mini-swe-agent

Contribute to this guide

Comments

Building a minimal AI agent from scratch

for software engineering, terminal use, and more

Authors: Kilian Lieret, Carlos Jimenez, John Yang, Ofir Press.

Contributors: Cesar Garcia, piot

( Contribute to this guide )

So you want to build your own AI agent from scratch? The good news: It's super simple, especially with more recent language models.<br>We won't be using any external packages (other than to query the LM), and our initial minimal agent is only some 60 lines long.

And if you think this is too simplified and can never work in practice: Our mini agent is built exactly the same, and is used for research at Princeton, Stanford, NVIDIA, Anyscale, essentials.ai and more.<br>Using this simple guide you can score up to 74% on SWE-bench verified, only a few percent below highly optimized agents.

Our first prototype in 50 lines&para;

Let's get started: From a top level view, an AI agent is just a big loop: You start with a prompt, the agent proposes an action, you execute the action, tell the LM the output, and then repeat.<br>To keep track of what have happend we continue to append to the messages list.

Pseudocode:

messages = [{"role": "user", "content": "Help me fix the ValueError in main.py"}]<br>while True:<br>lm_output = query_lm(messages)<br>print("LM output", output)<br>messages.append({"role": "assistant", "content": lm_output}) # remember what the LM said<br>action = parse_action(lm_output) # separate the action from output<br>print("Action", action)<br>if action == "exit":<br>break<br>output = execute_action(action)<br>print("Output", output)<br>messages.append({"role": "user", "content": output}) # send command output back

What's up with the role field?<br>The role field indicates who sent the message in the conversation. Common roles are:

"user" - Messages from the user/human

"assistant" - Messages from the AI model

"system" - System prompts that set context/instructions

Different LM APIs may have slightly different conventions for how to structure these messages.

So to get this to work, we only need to implement three things:

Querying the LM API (this can get annoying if you want to support all LMs, or want detailed cost information, but is very simple if you already know which model you want)

Parsing the action (parse_action). You don't need this if you use the tool calling functionality of your LM if it supports it, but this is more provider-specific, so we wo'nt cover it in this guide for now (don't worry, the performance should not be impacted by this).

Executing the action (very simple, in our case we will simply execute any action of the LM as a bash-command in the terminal).

Querying the LM&para;

Let's start with the first step. Click on the tabs to find the right LM for you.<br>litellm supports most specific LMs, so this is a good default option if your LM is not explicitly mentioned.

OpenAIAnthropicOpenRouterLiteLLMGLM

Install the OpenAI package (docs):<br>pip install openai

Here's the minimal code to query the API:<br>from openai import OpenAI

client = OpenAI(<br>api_key="your-api-key-here"<br>) # or set OPENAI_API_KEY env var

def query_lm(messages):<br>response = client.responses.create(<br>model="gpt-5.1",<br>input=messages<br>return response.output_text

Install the Anthropic package (docs):<br>pip install anthropic

Here's the minimal code to query the API:<br>from anthropic import Anthropic

client = Anthropic(<br>api_key="your-api-key-here"<br>) # or set ANTHROPIC_API_KEY env var

def query_lm(messages):<br>response = client.messages.create(<br>model="claude-sonnet-4.5",<br>max_tokens=4096,<br>messages=messages<br>return response.content[0].text

Install the OpenAI package (OpenRouter docs) - OpenRouter uses OpenAI-compatible API:<br>pip install openai

Here's the minimal code to query the API:<br>from openai import OpenAI

client = OpenAI(<br>base_url="https://openrouter.ai/api/v1",<br>api_key="your-api-key-here"<br>) # or set OPENROUTER_API_KEY env var

def query_lm(messages):<br>response = client.chat.completions.create(<br>model="anthropic/claude-3.5-sonnet", # or any model on OpenRouter<br>messages=messages<br>return response.choices[0].message.content

Install the LiteLLM package (docs) - supports 100+ LLM providers:<br>pip install litellm

Here's the minimal code to query the API:<br>from litellm import completion

def query_lm(messages: list[dict[str, str]]) -> str:<br>response = completion(<br>model="openai/gpt-5.1", # can be any provider + model<br>messages=messages<br>return response.choices[0].message.content

Install the Zhipu AI package (docs):<br>pip install zhipuai

Here's the minimal code to query the API:<br>from zhipuai import ZhipuAI

client = ZhipuAI(<br>api_key="your-api-key-here"<br>) # or set ZHIPUAI_API_KEY env var

def query_lm(messages):<br>response = client.chat.completions.create(<br>model="glm-4-plus",<br>messages=messages<br>return...