Context Course

0. Welcome to the Context Course

1. Skills: Portable Knowledge for Code Agents

2. MCP: The Model Context Protocol

3. Plugins: Bundling Tools for Distribution

4. Subagents: Orchestrating Multi-Agent Workflows

Unit 4: Subagents Subagent Patterns Using Subagents Quiz 1: Subagent Concepts Hands-On: Multi-Agent Workflow Quiz 2: Multi-Agent Workflows

5. Hooks: Observing and Guarding the Agent Lifecycle

6. Bonus: Nano Harness — Build Your Own Agent

Join the Hugging Face community

and get access to the augmented documentation experience

Collaborate on models, datasets and Spaces

Faster examples with accelerated inference

Switch between documentation themes

to get started

Quiz 2: Multi-Agent Workflows

Test your ability to design and implement multi-agent workflows.

Question 1: When is the overhead of subagents worth it?

Always spawn as many subagents as possible for maximum parallelism Use subagents when tasks are independent and parallelizable, and the overhead is justified by the task complexity Subagents are only for research tasks; never use for implementation Subagents should always share the same context window as the parent

Question 2: How do parent agents share context with subagents?

Subagents can directly access the parent agent's memory and variables Pass small context in the task description; for large data, write to a file and reference the path Use global variables that all subagents can read There's no way to share data between parent and subagent

Question 3: How should you design for subagent failure?

If a subagent fails, the entire workflow always fails Design for failure: use timeouts, handle partial results, and have fallback strategies Assume all subagents will succeed; don't add error handling since it's just overhead If a subagent fails, immediately spawn 5 more to retry the same task

Question 4: How do Claude Code and Codex expose subagents?

Claude Code: use /agent spawn. Codex: use codex-agent spawn. They're identical. Claude Code: conversational invocation + custom agents in .claude/agents/. Codex: natural language requests + custom agents as TOML files in .codex/agents/. Claude Code doesn't support subagents. Only Codex does. Both platforms use the exact same subagent API and configuration format

Question 5: What built-in agent types does Codex provide?

Codex has three built-in agent types: default (general), worker (implementation), and explorer (read-heavy codebase exploration) Codex has no built-in agent types; you must always define custom agents Codex only supports one agent at a time; there's no parallelism Codex subagents require a separate codex-agent binary to manage

Summary

If you got 4-5 correct, you’re ready to design multi-agent workflows without reaching for subagents unnecessarily. If you missed several, review the platform-specific invocation patterns and failure-handling guidance.

Key Takeaways

Use subagents when the task shape justifies isolation or parallelism
Pass only the context a child agent needs, and use files when the payload is too large for a prompt
Design for retries, partial results, and platform-specific invocation details from the start

Next Steps

You’ve finished Unit 4: Subagents. Next up is Unit 5, where you’ll add deterministic lifecycle hooks around those same agent workflows.

Update on GitHub

←Hands-On: Multi-Agent Workflow Unit 5: Hooks→

Quiz 2: Multi-Agent Workflows Question 1: When is the overhead of subagents worth it?Question 2: How do parent agents share context with subagents?Question 3: How should you design for subagent failure?Question 4: How do Claude Code and Codex expose subagents?Question 5: What built-in agent types does Codex provide?Summary Key Takeaways Next Steps