Build First AI Agent
Building reveals infinite loops, bad args, and context overflow that reading cannot teach.
Hello World for agents — small, ugly, foundation for everything after.
Visual Workflows
Start here — scroll inside each diagram frame to explore, then use + / − to zoom up to 200% if needed.
Scroll inside the frame to explore · use + / − to zoom up to 200%
Minimal Agent Loop
Scroll inside the frame to explore · use + / − to zoom up to 200%
The while-loop every framework implements under the hood.
File Layout
Scroll inside the frame to explore · use + / − to zoom up to 200%
Recommended project structure for your first no-framework agent.
Key Takeaways
- 1.Minimal agent: one LLM two tools while-loop logs.
- 2.Steps: goal prompt schemas loop max steps.
- 3.Learn failure modes before using frameworks.
- 4.Completable in one afternoon with OpenAI API.
Real Example
Scenario
Build an 80-line Python agent with `web_search` and `calculator` tools that answers: 'What is Japan's GDP per capita times 2?'
What you would do
Loop: user goal → LLM with tool schemas → if tool_calls, execute and append observation → repeat until text answer or max 10 steps. Log every iteration. Test on 5 questions mixing search and math. Monitor: steps used and tool selection accuracy.
Practice Task
Build the 2-tool agent with max 10 steps. Log every iteration to the console. Run it on 5 queries (mix search + math) and note where it fails.
Code Walkthrough
Highlighted lines show where Build First AI Agent happens in the code.
1from openai import OpenAI # import dependencies2
3client = OpenAI() # create API client4tools = [5 {"type": "function", "function": {"name": "web_search", "parameters": {"type": "object", "properties": {"query": {"type": "string"}}}}},6 {"type": "function", "function": {"name": "calculator", "parameters": {"type": "object", "properties": {"expression": {"type": "string"}}}}},7]8
9def run_agent(goal: str, max_steps: int = 10) -> str: # define a reusable function10 messages = [{"role": "user", "content": goal}]11 for step in range(max_steps):12 resp = client.chat.completions.create(model="gpt-4o-mini", messages=messages, tools=tools) # core API call for Build First AI Agent13 msg = resp.choices[0].message14 print(f"Step {step + 1}:", msg.tool_calls or msg.content) # show output for debugging15 if not msg.tool_calls:16 return msg.content or "" # return the result17 messages.append(msg)18 # execute tool, append {"role": "tool", ...}, continue loop19 return "max steps reached" # return the result20
21print(run_agent("Japan GDP per capita times 2 — show sources")) # main loop: LLM → tools → append → repeatCheat Sheet
Quick recap
quick ref- •2-3 tools max at first
- •max_steps=10 always
- •Log every iteration
- •Validate tool JSON args
- •No framework until loop works
- •API key in .env never in code
Common Mistakes
- ✕Skipping evaluation for Build First AI Agent before production
- ✕No logging or tracing around build first ai agent steps
- ✕Ignoring cost and latency implications