What is an AI Agent?
LLMs alone only generate text. Agents complete multi-step workflows by calling APIs, databases, and external systems.
A chatbot advises from a chair; an agent logs in, pulls data, drafts the report, and sends it for review.
Visual Workflows
Start here — scroll inside each diagram frame to explore, then use + / − to zoom up to 200% if needed.
Overview
Scroll inside the frame to explore · use + / − to zoom up to 200%
Scroll inside the frame to explore · use + / − to zoom up to 200%
End-to-End Agent Runtime
Scroll inside the frame to explore · use + / − to zoom up to 200%
User goal through bounded loop to final output — with guardrails and HITL.
Five Components Working Together
Scroll inside the frame to explore · use + / − to zoom up to 200%
How brain tools memory planning and loop connect in one run.
Key Takeaways
- 1.Agent = LLM reasoning + tools + memory + loop until goal is done.
- 2.Chatbots answer once; agents observe, act, and iterate.
- 3.Five components: brain, tools, memory, planning, orchestration.
- 4.Production needs limits, guardrails, tracing, and human approval.
Real Example
Scenario
A sales VP asks: 'Summarize Q3 revenue by region and email the board.' The agent must query Snowflake, build a chart, draft the email, wait for approval, then send via Gmail.
What you would do
Break the goal into subtasks: (1) call `run_sql` on Snowflake for Q3 regional revenue, (2) pass rows to `create_chart`, (3) draft the email with the chart attached, (4) pause at a human-approval gate because `send_email` is destructive, (5) send and return the message ID. Monitor: tool success rate, steps per task, and approval bypass attempts.
Practice Task
Pick one recurring task from your work (weekly report, ticket triage, data pull). Write 5–7 agent loop steps, name the tool for each step, and mark where human approval is required.
Code Walkthrough
Highlighted lines show where What is an AI Agent? happens in the code.
1from openai import OpenAI # import dependencies2
3client = OpenAI() # create API client4tools = [5 {"type": "function", "function": {"name": "run_sql", "parameters": {"type": "object", "properties": {"query": {"type": "string"}}}}},6 {"type": "function", "function": {"name": "send_email", "parameters": {"type": "object", "properties": {"to": {"type": "string"}, "body": {"type": "string"}}}}},7]8
9messages = [{"role": "user", "content": "Summarize Q3 revenue by region and email the board"}]10MAX_STEPS = 811
12for step in range(MAX_STEPS):13 response = client.chat.completions.create( # core API call for What is an AI Agent?14 model="gpt-4o-mini",15 messages=messages,16 tools=tools,17 )18 msg = response.choices[0].message19 if not msg.tool_calls:20 print("Final:", msg.content) # show output for debugging21 break22 # Execute tool, append result, loop again23 messages.append(msg)24 messages.append({"role": "tool", "tool_call_id": msg.tool_calls[0].id, "content": '{"rows": [...]}'})Cheat Sheet
Quick recap
quick ref- •Agent = LLM + Tools + Memory + Loop
- •Observe → Reason → Act → Update → Repeat
- •max_iterations + cost budget required
- •Trace every step for debugging
- •Human approval for writes and deletes
- •Sandbox tool execution in production
Common Mistakes
- ✕No iteration limit on the agent loop
- ✕Unrestricted tool access without sandboxing
- ✕Treating agents as single-turn chatbots