Subagents

• Key points are in the visual diagram above.

One giant agent with every tool pollutes context and over-permissions.

A tech lead assigning a well-scoped ticket, not handing over root.

Scenario

A tech lead assigning a well-scoped ticket, not handing over root.

Solution

In Claude Agent SDK, apply Subagents to this scenario: A tech lead assigning a well-scoped ticket, not handing over root. Identify the inputs, run the technique, validate the output, and note one thing you would monitor in production.

Do this before moving to the next module — reading alone is not enough.

Open the Code Walkthrough below and run it locally. Change one parameter related to Subagents (e.g. model, temperature, top_k, or tool name), observe the difference in output, and write 2–3 sentences explaining what changed.

Highlighted lines show where Subagents happens in the code.

1# Subagents — minimal example2from openai import OpenAI3
4client = OpenAI()  # create API client5
6# Ask the model to explain this topic7response = client.chat.completions.create(  # core API call for Subagents8    model="gpt-4o-mini",9    messages=[10        {"role": "system", "content": "You explain subagents clearly."},11        {"role": "user", "content": f"What is subagents?"},12    ],13    temperature=0,14)15print(response.choices[0].message.content)  # show output for debugging

• Skipping evaluation for Subagents before production
• No logging or tracing around claude subagents steps
• Ignoring cost and latency implications

Quick recap — the most important points from this module.