Level 8 — Delegate: Subagents and Parallelism

10 questions, drawn from a bank of 35 across 28 topics. Pass at 80%. Every wrong option carries a note explaining why it is tempting — those notes are the point of the quiz, so read them whether you pass or not.

attempt 1

1 / 10 · Choose one
Claude Code opens with a startup notice that your agent descriptions are over the 15.0k-token limit. You have a dozen custom agents and have delegated to none of them yet. What is that notice telling you?
Claude Code opens with a startup notice that your agent descriptions are over the 15.0k-token limit. You have a dozen custom agents and have delegated to none of them yet. What is that notice telling you?
Claude Code stopped loading agents once the running total crossed 15,000 tokens, so the definitions it had not reached yet are not available this sessionNothing is spent until Claude routes work to one of them, so a session where you never delegate pays none of that totalThe system prompts in the bodies of those files are what pushed the total over, so shortening the instructions inside each file is the fixEvery agent loaded anyway — the total is a notice, not a cap — and its cost lands in your window whether or not any agent runs
2 / 10 · Choose one
In the documented walkthrough a subagent reads 6,100 tokens of files to answer a research question. What lands in your context window?
In the documented walkthrough a subagent reads 6,100 tokens of files to answer a research question. What lands in your context window?
Nothing until you ask for it — the result is held in the task list and only pasted in on requestA compressed version of every file it read, so the summary can be checked against its sourcesThe 420-token summary and the file reads that produced it, so you can audit the reasoningIts final text response, plus a small metadata trailer with token counts and duration
3 / 10 · Choose one
The documented code-reviewer example is described as read-only. What does that list buy you that an instruction in the prompt would not?
terminal
```
---
name: code-reviewer
description: Expert code review specialist. Proactively reviews code for quality, security, and maintainability. Use immediately after writing or modifying code.
tools: Read, Grep, Glob, Bash
model: inherit
---
```
The documented code-reviewer example is described as read-only. What does that list buy you that an instruction in the prompt would not?
The list decides what the reviewer can do, not what it will do: with Edit and Write absent a finding cannot become a fixIt means any edit the reviewer attempts raises a permission prompt in your session that you can declineIt is belt and braces — the body already tells the reviewer to report and not to edit, and that instruction is the real controlFewer tools means a shorter tool listing in its context, so the review is cheaper and faster
4 / 10 · Choose one
With agent view on, so /subtask and /fork are distinct commands, you type /fork draft the migration notes and keep working. Nothing comes back. Why?
With agent view on, so /subtask and /fork are distinct commands, you type /fork draft the migration notes and keep working. Nothing comes back. Why?
You were moved into the copy, so the notes are being written in the conversation you are no longer looking atThe result is waiting behind a permission prompt in the copy that nobody has answered yet/fork copied the session into a new background session, and nothing the copy does enters the original on its ownIt is simply still running — a forked subagent reports back only when it finishes, and drafting notes takes a while
5 / 10 · Choose one
A colleague puts the reviewer's real instructions in its description and leaves the body of the file as two lines of notes about what the agent is for. Why does it behave nothing like they expected?
A colleague puts the reviewer's real instructions in its description and leaves the body of the file as two lines of notes about what the agent is for. Why does it behave nothing like they expected?
The description is only read at startup, so instructions placed in it arrive too late to shape any behaviourThe body is the system prompt, so those two lines of notes are the whole brief the subagent runs onFrontmatter is configuration, so a description past one line is truncated before the definition loadsThe Claude Code system prompt loads underneath the body and overrides two lines of notes
6 / 10 · Choose one
Your subagent returns an empty, confusing result. Its disallowedTools happens to remove every entry in its tools list. What did Claude Code do?
Your subagent returns an empty, confusing result. Its disallowedTools happens to remove every entry in its tools list. What did Claude Code do?
Launched it without tools — the zero-tools refusal is skipped when disallowedTools empties the listRefused to launch it, returning an error that names each entry it could not resolve to a toolFell back to the inherited pool, since a definition with no usable tools is treated as if tools had been omittedSkipped the agent file entirely and wrote the tool-list conflict to the debug log instead
7 / 10 · Choose one
Given this frontmatter, which tools does the subagent end up with, and by what route?
terminal
```
---
name: auditor
description: Audits a module and reports findings
tools: Read, Bash
disallowedTools: Bash, Write
---
```
Given this frontmatter, which tools does the subagent end up with, and by what route?
Read and Bash — an explicit tools allowlist is more specific than a denylist, so it wins for a tool named in bothRead and Bash, with Write denied — disallowedTools only subtracts from the inherited pool, so it cannot touch a tool you listed explicitlyRead only — disallowedTools is applied first, then tools is resolved against what is left, and a tool listed in both is removedNone — listing Bash in both fields is a contradiction, and Claude Code refuses to load the definition
8 / 10 · Choose one
A colleague says: "delegate it to Explore — it is already reading those files, so let it fix the config while it is in there." Why can that not happen?
A colleague says: "delegate it to Explore — it is already reading those files, so let it fix the config while it is in there." Why can that not happen?
It could, but each edit would raise a permission prompt in your main session firstExplore is a read-only agent: Write and Edit are denied to it, so the fix comes back as a findingExplore skips your CLAUDE.md, so it would not know the project conventions well enough to edit safelyExplore only runs inside plan mode, where the whole session is read-only and nothing may write
9 / 10 · Choose one
You @-mention a specific subagent and write a careful four-paragraph brief. What actually reaches the subagent?
You @-mention a specific subagent and write a careful four-paragraph brief. What actually reaches the subagent?
A task prompt Claude composed from your message — the @-mention picks the agent, not the promptNothing at first — the subagent asks the main conversation for context when it needs itOnly the agent's own description field, which is the text Claude routes on and hands acrossYour four paragraphs verbatim, since naming the agent hands the message straight to it
10 / 10 · Choose all that apply
A custom subagent starts. Select everything that is in its window before it does any work of its own.
A custom subagent starts. Select everything that is in its window before it does any work of its own.
The files Claude already read in your conversationA git status snapshot taken at the start of the parent sessionThe CLAUDE.md hierarchy the main conversation loadsA condensed transcript of your session so far, for backgroundIts own system prompt — the agent file body plus environment details, not the Claude Code system promptA task message: the delegation prompt Claude wrote when it handed off the work

10 questions still to answer