What is an AI agent? How it differs from a chatbot and a workflow

Before you start

Nothing to install. Two ideas from earlier lessons carry over: a system prompt tells an agent how to behave, and the context window is the room it works in. If you have Claude Code you can watch a real agent loop, but reading is enough.

Tell an agent from a chatbot with three questions

Product names will not settle it, so ask these questions in order. The sketch below is ours. It follows OpenAI's description of the Agents SDK loop, which runs until a final-output tool is called or the model answers without a tool call.

messages = [instructions, user_request]
while True:
    reply = ask_model(messages, tools)
    if reply has no tool call:
        return reply              # a plain answer ends the run
    result = run_tool(reply.tool_call)
    messages += [reply, result]   # the result feeds the next step
    # also stop when a step limit is reached

1. Who picks the next step? In a workflow, your code does. Anthropic's article defines workflows as systems where models and tools are orchestrated through predefined code paths, and agents as systems where the model dynamically directs its own process and tool use.
2. What can it touch outside the chat? OpenAI says an agent has tools to interact with external systems, both to gather context and to take actions. A system that only returns text is not doing that.
3. How does it know it is done? At each step it checks what came back from the environment, which Anthropic calls "ground truth", such as a tool result. The run ends when the task is complete or a limit is hit.

Chatbot, workflow or agent

Judge a product by the three questions, not by its label, since many chat products now add tools and drift toward the agent column. The table uses Anthropic's split between workflows and agents, plus OpenAI's remark that simple chatbots, single-turn models and sentiment classifiers are not agents.

Model, tools and instructions

OpenAI's guide says an agent has three core components. The model does the reasoning and decision-making, the tools are functions or APIs it can call to act, and the instructions are the explicit guidelines and guardrails for its behavior. The instructions are a system prompt.

The guide sorts tools into three types. Data tools retrieve context. Action tools change something, such as sending an email or updating a record. In orchestration tools, one agent serves as a tool for another. Tools can come from function calling or from MCP, and OpenAI's agents guide lists both. The guide's own example is an agent with instructions and two tools.

from agents import Agent, WebSearchTool, function_tool

@function_tool
def save_results(output):
    db.insert({"output": output, "timestamp": datetime.time()})
    return "File saved"

search_agent = Agent(
    name="Search agent",
    instructions="Help the user search the internet and save results if asked.",
    tools=[WebSearchTool(), save_results],
)

A run in Claude Code

Claude Code's docs describe the loop as three phases that blend together: gather context, take action and verify results, with tools throughout. Without tools Claude can only respond with text, and with them it can act. For a request to fix the failing tests, the docs list what Claude might do. You can interrupt at any point to steer it.

When an agent is worth building

Anthropic's advice is to find the simplest solution possible and add complexity only when needed. Agents suit open-ended problems where you cannot predict the steps, but their autonomy means higher costs and the potential for compounding errors.

OpenAI gives three signs that a job has resisted ordinary automation. If none of them fits, its guide says a deterministic solution may be enough.

Limits and a person in the loop

• Set a stop condition. Anthropic suggests a maximum number of iterations, and OpenAI lists exits such as a final-output tool, an error or a maximum number of turns.
• Rate each tool by risk. OpenAI suggests low, medium or high, based on read-only versus write access, reversibility, account permissions and financial impact, then pausing for checks or a human on the high ones.
• Hand control to a person on two triggers. OpenAI names exceeding failure thresholds and high-risk actions, such as canceling orders or making payments.

Ethan Mollick on why running agents is a management job

Ethan Mollick, a Wharton professor who studies AI, wrote on X on 3 January 2026 that watching people use coding agents such as Claude Code and Codex makes one thing clear. Managing agents is a management problem. His four questions are whether you can specify goals, provide context, divide up tasks and give feedback, and he calls them teachable skills.

Each one maps onto a lesson in this course. Goals and limits go in the system prompt, and context is what you choose to put in the window. Dividing up tasks is the workflow-or-agent choice above, and feedback is the person in the loop on risky steps. Before you hand a job to an agent, write one line for each of his four questions.

A screenshot of X's own embed of the post, taken on 5 November 2026, with the time in UTC. Select it to open the post on X.

Common mistakes

• Reaching for an agent when a fixed workflow would do. Anthropic notes that optimizing a single call with retrieval and examples is often enough.
• Starting with several agents. OpenAI says to maximize one agent's capabilities first, because more agents add complexity and overhead.
• Giving an agent many overlapping tools. OpenAI says the problem is not only the count, since some setups handle more than 15 distinct tools while others struggle with fewer than 10 overlapping ones.
• Skimping on tool descriptions. Anthropic says tool definitions deserve as much prompt engineering effort as the main prompts.

Try it yourself

• Pick a job you do by hand and write down its steps. Mark each one as something a fixed rule could handle or something that needs judgment.
• Run the three questions on an AI product you use, and decide whether it is a chatbot, a workflow or an agent.
• If you use Claude Code, ask it a question about your project and watch which tools it calls. Name the phase of each step.

Where to go next

That ends this course. Next is which AI tool to use. Its first lesson, how to choose an AI image generator, compares the apps that make pictures and says how to pick one for a job. The course then does the same for video, voice, music, research and AI agents.

The Learn hub lists every course. Mindket also lists ready-made agents, so you can see one before you build your own.

Frequently asked questions

Sources

1. Anthropic Engineering: Building effective agents (read November 4, 2026)https://www.anthropic.com/engineering/building-effective-agents
2. OpenAI: A practical guide to building agents (read November 4, 2026)https://cdn.openai.com/business-guides-and-resources/a-practical-guide-to-building-agents.pdf
3. OpenAI: Agents guide (read November 5, 2026)https://developers.openai.com/api/docs/guides/agents
4. Claude Code docs: How Claude Code works (read November 5, 2026)https://code.claude.com/docs/en/how-claude-code-works
5. Ethan Mollick on X, 3 January 2026 (read November 5, 2026)https://x.com/emollick/status/2007249835465072857
6. YouTube: Anthropic, Tips for building AI agents (read November 5, 2026)https://www.youtube.com/watch?v=LP5OCa20Zpg