Building agents and MCP servers

In Module 07 you watched the agent loop from the outside. Now you write it: about twenty lines of code, a few well-described tools, and the limits that keep it safe. Then you turn your own system into an MCP server, so every agent from Module 10 can use it.

1. The idea

In code, those four parts are:

1. Brain: a model API call with a clear system prompt (Modules 05 and 14).
2. Hands: tools, each a name, a description and an input schema, backed by your function (Module 06).
3. Loop: call the model; if it asks for tools, run them and send the results back; repeat until it answers.
4. Leash: a maximum number of turns, a token or cost budget, read-only tools by default, approval for risky actions, and logs of every step.

Start with the simplest thing that works. Anthropic's advice in Building effective agents is to use a single call or a fixed workflow when the steps are predictable, and a full agent only when they aren't.

2. How it works

Here is the whole loop with the Claude API. The other providers' versions have the same shape.

import anthropic

client = anthropic.Anthropic()                       # reads ANTHROPIC_API_KEY

def run_agent(task, max_turns=10):
    messages = [{"role": "user", "content": task}]
    for turn in range(max_turns):                    # leash: turn budget
        reply = client.messages.create(
            model="claude-sonnet-5-5", max_tokens=2048,
            system=SYSTEM_PROMPT, tools=TOOLS, messages=messages,
        )
        messages.append({"role": "assistant", "content": reply.content})
        if reply.stop_reason != "tool_use":          # no tool requested: done
            return next(b.text for b in reply.content if b.type == "text")
        results = []
        for block in reply.content:
            if block.type == "tool_use":
                output = run_tool(block.name, block.input)   # your code, your checks
                results.append({"type": "tool_result", "tool_use_id": block.id,
                                "content": output})
        messages.append({"role": "user", "content": results})
    raise RuntimeError("Turn budget exhausted")      # stop, don't loop forever

3. Try it yourself

The leash: turn budgets

A task needs a certain number of tool calls. Some runs go wrong and repeat themselves. See what the budget does.

Write an MCP server tool

Fill in one tool and get a working MCP server in Python (official SDK, v2). Any MCP client, including all five coding tools from Module 10, can then call it.

4. Why it matters at work

1. Design tools for the model, not for your API. Fewer, task-shaped tools ("get_order_status") beat thin wrappers over every endpoint. Clear names, descriptions with examples, and short, useful outputs.
2. Return errors the model can act on. "Order A-1042 not found; order numbers look like A-1234" lets it recover. A stack trace doesn't.
3. Read-only first. Add tools that change things only when needed, and put approval in front of anything irreversible (sending, paying, deleting).
4. Treat tool output as untrusted. A web page or ticket can contain instructions aimed at your agent (prompt injection, Module 18). Your code, not the prompt, enforces permissions.
5. Log every turn: the model's request, the tool input and output, tokens and time. You'll need it to debug and to evaluate (Modules 17 and 19).
6. Use an SDK once the basics are clear. SDKs give you the loop plus context management, sessions, subagents, tracing and permission handling.

5. Agent SDKs and protocols

Checked against vendor documentation, October 2026.

MCP and A2A: tools vs teammates

6. Remember it

7. Common mistakes

8. Hands-on lab (roadmap week 18)

1. Write the loop yourself with two read-only tools (for example: search notes, read a note). Cap it at 8 turns and log every step.
2. Break it on purpose: a tool that errors, a vague tool description, a task that can't be done. Watch how the agent reacts and improve the tool messages.
3. Rebuild it with an SDK (Claude Agent SDK or OpenAI Agents SDK) and compare: what did the SDK add for free?
4. Build an MCP server exposing one tool from your own system, test it with the MCP Inspector, then connect it to your coding tool.
5. Add the leash: one tool that changes something, gated by an approval step.

9. Self-check quiz

10. Go deeper