What actually separates an Agent from a Chatbot?

A Chatbot only talks — single turn, fixed flow. An Agent acts — multi-step planning, autonomous decisions. Tell a Chatbot "analyze this company's financials" and it lists the steps; tell an Agent the same and it searches the filings → extracts metrics → benchmarks against the industry → writes the report. Three things together make it an Agent: callable tools (search, code exec, DB, API) + task decomposition + short-term and long-term memory.

What's the difference between tool_choice="auto" and a manual choice in OpenAI Function Calling?

`tool_choice="auto"` lets the model decide "do I need a tool, and which one"; you can pass `{"type": "function", "function": {"name": "search_web"}}` to force a specific call. The chapter demo uses auto + gpt-5 + a single search_web tool (param: query) — given "what's the weather today?", the model decides on its own to fetch from the web. That's Agent decision-making in its smallest form. In production, high-risk steps (like committing code) skip auto — manual dispatch or human approval instead.

What goes into Agent short-term vs long-term memory?

Short-term memory = current conversation context (recent turns, tools just called, key results). Long-term memory = persistent knowledge and preferences (user's past questions, enterprise knowledge base, preferred output format). Short-term lives in an in-memory list or ConversationBufferMemory; long-term usually sits in a vector store + key-value store. The classic beginner mistake: cramming everything into short-term — context fills up, lost-in-middle kicks in, and important info gets forgotten.

How does an Agent decompose "analyze a company's finances" into executable steps?

Through planning. Given "analyze this company's finances," the Agent internally decomposes into: (1) search filings; (2) extract key metrics (revenue, margin, debt ratio); (3) benchmark against industry averages; (4) write the report. Each step maps to a tool call + result writeback. Planning quality depends on whether the system prompt provides examples, the tool list is clear, and the model is asked to output a plan before executing — ReAct (Reasoning + Acting interleaved) is the standard pattern.

What capabilities does a research assistant Agent need at minimum?

Three things: (1) web search (search_web tool, param: query); (2) document reading and analysis (fetch_url or read_pdf, plus OCR/chunking); (3) report generation (structured output, ideally section-by-section). Underneath: a Function Calling loop — model picks the next tool → executes → feeds result back → continues until the report is done. For production, add: max_steps (kill infinite loops), timeout, mandatory citations (every claim tied to a source). That's the gap between practice toy and shippable agent.

AI Agent Development Basics

⏱️ 90 min

Build autonomous AI agents using the ReAct pattern

🔒

Register to Unlock Full Content

✓Full chapter content
✓Video tutorials
✓Code examples
✓Learning community