Claude Code Without Coding · Section 10

Notes: Agents and Subagents

The ideas worth keeping from Section 10, written to read instead of watch: what an agent receives, how it works and what you need back.

Arun Nagarathanam · 12 Oct 2026 Notes · 72 of 80 in this course

Lecture 10.1

What Does a Subagent Do?

I can hand one part of a job to a separate agent and keep my main conversation on the decision I actually need to make.

An agent does not see your whole conversation. It gets the assignment and the starting information Claude hands it.

Key ideas

01Main agent: My main conversation. It directs the work, writes the assignment and uses what comes back.
02Subagent: An agent working on the one part it received, in its own working conversation.
03Built-in agents: Claude Code already has some. Claude picks one from the job, or I name it.
04Saved agent: When the same responsibility keeps coming back, I save it with its own instructions.

Across this section we follow Willow Desk, a fictional business with supplied files. I will:

question a plan with one agent
save that agent so I can reuse it
run 3 agents on 3 different files at once
research a local product page before changing it
adapt an agent someone else wrote

The section in one picture: one agent first, then 3 agents reading 3 different files while the main agent keeps the decision.

What stays with me. I decide what useful work to assign and how to judge what comes back.

Lecture 10.2

Give an Agent One Useful Job

A plan can sound sensible and still leave a question open. Giving an agent a perspective and one clear job is how I find that question before I commit.

The report counted 3 colour comments, but F06 asks for a second colour without naming one, so only 2 support the proposed colours.

Key ideas

01Perspective: I asked for a first-time customer's view, so the review focused on what a buyer needs to understand.
02One job: Examine the plan, connect concerns to the supplied comments and return questions.
03Check the evidence: An agent can stretch a comment. I open the cited IDs before I believe a finding.

My request named 4 things:

the 2 files: business-plan.md and customer-feedback.md
the perspective: a first-time customer
the return: up to 5 concerns, each with a next question
the limits: leave the files unchanged and keep the decision with me

I also asked the main conversation to save the critique as initial-plan-critique.md and tell me which agent ran.

The useful finding. The plan reuses the existing setup guide, and several comments say that guide is hard to find or follow.

The stretched finding. The report counted 3 colour comments as support for the pilot. One of them asks for a second colour without naming it, so it supports neither proposed colour.

Make it yours

Freelancer or consultant

Your input
A proposal I am about to send a client

Ask Claude
Critique it as the client would read it, up to 5 questions

Check
Each question points to a real line in the proposal
Small business owner

Your input
A plan for a new product or offer, plus recent customer comments

Ask Claude
Review it from a first-time customer's view, tied to comment IDs

Check
Open each cited comment and confirm it says what the report claims
On a team at work

Your input
A project plan and the feedback it responds to

Ask Claude
Review it for what a reader might misunderstand

Check
The decision is still left to me, not made in the report

Lecture 10.3

What Should Your Agent Receive and Return?

An agent only knows what Claude gives it. When I open the assignment Claude sent, I can check that my meaning survived the handoff.

A complete handoff names the purpose, the starting files, the actions allowed and the result needed back.

Key ideas

01Context: The information a model can work with during a conversation.
02Separate context: The agent has its own conversation. It does not mean another computer or a copy of every file.
03Handoff: Purpose, starting information, allowed actions and the result needed back.
04The return: It takes space in my main conversation, so I ask for a short return and keep detail in a named file.

The prompt the main agent sent was longer than what I typed. It named both files, the first-time customer's view, what I wanted back, and told the agent to leave the files unchanged.

A handoff goes wrong in two easy ways:

Only a filename. The agent cannot tell if I want a summary, a critique or a rewrite.
A missing file. I describe the critique but leave out the feedback, so the agent has no comments to use.

A token count tells me how much work the model processed. On its own it does not tell me how much room is left in my main conversation.

Lecture 10.4

Create an Agent You Can Reuse

When a review is worth repeating, I save it as a named agent. Next time I only supply the files.

The agent file: opening fields identify the agent and set its tools, model and effort, and the body explains the work.

Key ideas

01Agent file: A Markdown file in .claude/agents/. Opening fields identify it; the body explains the work.
02Description: Tells Claude when the agent is useful, so a clear one helps Claude select it.
03Tools: customer-plan-critic lists only Read, which is enough when I name the files.
04Scope: Project agents live with the project. Personal agents in ~/.claude/agents/ work across projects.

In customer-plan-critic.md I checked:

name: customer-plan-critic, the name I use in a prompt
description: a first-time-customer review of a plan and feedback
model: Sonnet, with medium effort
tools: Read
instructions: a table of concern, evidence and next question, then missing information

When it is not found. My first call said the agent was not registered. I had created the agents folder during that session, so I started a new conversation in the same project and it ran.

If a named agent is missing, check whether the session loaded its file before blaming the instructions.

Make it yours

Freelancer or consultant

Your input
A review I give every client draft

Ask Claude
Save it as a project agent with a clear name and description

Check
The tools field lists only what the review needs
Small business owner

Your input
A monthly check of a plan against customer comments

Ask Claude
Create a named agent for that check

Check
Call it by name in a new conversation and see that it runs
On a team at work

Your input
A review the whole team repeats

Ask Claude
Create it as a project agent so it travels with the project

Check
Its description says exactly when to use it

Lecture 10.5

Can a Skill Call Agents?

A Skill can hold instructions for several roles and have the main agent hand each role to a separate agent. That let me review one video from 3 angles at once.

The Skill held the instructions, 3 agents each took one file, and the main agent wrote the combined brief.

Key ideas

01Role instructions: Headings inside one Skill, not 3 saved agents. Claude gives each role to a general-purpose agent.
023 jobs: Audience reads 8 viewer comments, performance compares the figures, script reviews the written plan.
03Main agent: Checks the reports, ticks the shared checklist and writes the combined brief.

The 3 files for one 45-second setup video:

video-script.md: what the video says and shows, by time range
viewer-comments.csv: 8 viewer comments
performance.csv: views and average watch time for this video and 3 earlier ones, at 48 hours

Who ticks the checklist. An agent finishing its own review does not update the shared progress file. The main agent does, after it checks each report.

Both choices can work together:

save a Skill when you want to reuse instructions
save a named agent when the same responsibility keeps returning

Make it yours

Freelancer or consultant

Your input
A client deliverable plus their feedback and any numbers

Ask Claude
Review it with separate agents for feedback, figures and the draft

Check
Each recommendation names the comment or figure behind it
Small business owner

Your input
A video or post, its comments and its stats

Ask Claude
Run a review Skill with audience, performance and content roles

Check
The main agent marks a stage done only after reading its report
On a team at work

Your input
A launch asset reviewed by several people

Ask Claude
Give each review its own agent and combine the findings

Check
Disagreements between reports are shown, not hidden

Lecture 10.6

Give Different Agents Different Jobs

Agents can start together only when each already has what it needs. The step that combines their work has to wait for them.

The 3 reviews began together, and the main agent combined their reports only after all 3 returned.

Key ideas

01Parallel: Each agent had its own input and question, so none waited for another.
02Dependent step: The combined brief needed all 3 reports before it could start.
03Surfaced disagreement: Comment V05 asked for setup first; the script agent kept colours first for 7 seconds.

A recommendation with support. The brief says to name each part consistently, quoting a viewer asking which part "this piece" means. The script says "this piece" where the shot plan names the cable box.

The brief offered 2 openings:

start with the first setup step and show colours later
shorten the colour opening and begin setup sooner

I would open with the first setup step, because the goal is helping someone follow the setup. That is a proposed choice until I make the video and see how viewers respond.

What the numbers can say. The current video's average watch time is shorter than the 3 earlier videos. That is a reason to investigate. It does not explain why viewers stopped.

Lecture 10.7

Research a Project Before Changing It

I let a research agent read and search first, then I choose the change. My main conversation stays on the decision instead of on every source.

Claude proposed a broader change. I approved the design and 3 additions, then checked the revised guide in the browser.

Key ideas

01Research agent: Reads files and searches the web in its own conversation, then returns a brief.
02Local files: The pages were HTML on my computer. Nothing was published.
03localhost: A preview address served from my own computer.

What I did, in order:

tried the links and found that none reached their destination
asked a research agent for unanswered setup questions and 3 improvements, saved as setup-research-brief.md
chose one change: a short setup section and a clear guide link
described each failed click and asked Claude to wait for my approval on extra content
approved a parts list, a line saying when setup is done and a handling note

Separate problems stay separate. The brief kept damaged packaging apart from setup questions. A guide link helps someone find instructions, while a packaging complaint needs its own investigation.

Research gives me reasons. I still pick the change and check it in the browser.

Make it yours

Freelancer or consultant

Your input
A client's page and the questions their customers ask

Ask Claude
Research what the page leaves unanswered and suggest 3 improvements

Check
Each suggestion names the question it answers
Small business owner

Your input
My product page, setup guide and customer feedback

Ask Claude
Research first, save a brief, change nothing yet

Check
The original files are unchanged until I approve an edit
On a team at work

Your input
A help page or internal doc people keep asking about

Ask Claude
Find the gaps, then make one approved change

Check
I open the page and click every changed link

Lecture 10.8

Choose What an Agent Can Do

An agent's tools decide what it can do, and its instructions decide what it should do. I match both, and the model, to the job.

Each job gets its own tool and reasoning setting: the critic and performance agent read, and the page job edits.

Key ideas

01Tools field: Read lets the critic open files. Saving a report needs Write too.
02Model and effort: Saved per agent. Try a smaller model for simple retrieval; use a stronger one for hard judgment.
03maxTurns: An optional limit on rounds of work. It is not a cost limit or a schedule.
04Agent team: Agents that coordinate directly. My demos used subagents that report back instead.

When I asked the critic to save customer-plan-critique.md itself, Claude pointed out it had only Read. I replied "Yes, add Write and use that filename."

A changed file is not a working run. The definition now listed Read and Write, but that attempt still ran without Write and saved nothing. The next step is a fresh conversation, then opening the report.

Usage adds up:

every agent's work counts alongside the main conversation
3 agents rereading the same files for the same question add work, not value

Approving an action is separate from configuring a tool. Reading practice comments and publishing a live page carry different consequences.

Make it yours

Freelancer or consultant

Your input
An agent that should save its own report

Ask Claude
Add only the tool the job needs and name the file

Check
Run it in a fresh conversation and open the saved file
Small business owner

Your input
An agent doing a simple lookup

Ask Claude
Try a smaller model on it

Check
It still finds the right material
On a team at work

Your input
An agent near anything that publishes or sends

Ask Claude
Keep it read-only and return a draft

Check
Publishing still needs my explicit approval

Lecture 10.9

Find and Adapt Existing Agents

Someone may already have written an agent close to what I need. I treat it as a starting point and cut it down to my job.

An existing agent is the starting point. Our guide shapes it into the post we actually want.

Key ideas

01Community agents: VoltAgent's collection has files for content marketing, business analysis and user research.
02Narrow it: content-marketer covered strategy, campaigns and measurement. I kept the writing guidance only.
03Credit: Claude saved a notice with attribution and the MIT licence beside the adapted agent.

How the adaptation went:

saved setup-content-writer, then got "Agent type not found"
opened a new conversation and ran it by name
compared the draft with the setup guide
cut "saves guessing later", a benefit the guide never states
added a line saying Willow Desk is a fictional teaching example

I also had the main conversation start a separate agent for a 5-page Canva carousel. I opened the design but did not download the PDF.

Updated, not tested. I saved a carousel update to setup-content-writer. The next run has to produce both the design and the PDF before I trust it.

Make it yours

Freelancer or consultant

Your input
A community agent close to a service I offer

Ask Claude
Adapt it to one client job and remove the rest

Check
Every claim in the output appears in my source material
Small business owner

Your input
My own guide or product page

Ask Claude
Adapt a writing agent to turn it into one post draft

Check
Part names and steps match the guide
On a team at work

Your input
Meeting notes or customer comments

Ask Claude
Adapt business-analyst or ux-researcher to that material

Check
The licence and source stay beside the adapted file

Lecture 10.10

Which Job Deserves an Agent?

Most problems with agents come from 5 places. Knowing which one I have tells me which fix to reach for.

Problem 5: the main agent ticks the shared checklist as each return arrives, so you can see what finished.

Key ideas

01agent-fit: The course Skill that interviews me one question at a time: up to 5, then up to 5 more if needed.
024 outcomes: Keep it in the main conversation, give an existing agent one part, create a reusable agent, or refine the idea.
03Saved is not tested: plan-feedback-reviewer was created, but the named-agent test failed in that session.

The 5 problems and their fixes:

Choosing the work. Describe the result, then use agent-fit to find a useful split.
A misunderstood job. Say what you expected and why, then compare the return with the assignment, inputs and actions.
The wrong agent. Name the agent, see which one ran, and check its loaded description.
Agents without a useful split. Give each a distinct question, its inputs and its own report.
Not knowing what finished. Each agent says what it completed, where the result is and what remains.

A rule I would refine. The proposed brief said every claim should cite feedback IDs. A point about the plan itself should point to the plan instead.

Every fix has a matching prompt in FIVE-PRACTICAL-SOLUTIONS.md in the practice pack.

Lecture 10.11

Choose Your Next Delegation

Every job in this section started with the same decision: which part can an agent take, and what do I need back to use it?

Every assignment names a job, its inputs and a result. A review names questions and material, and a change names a target and the action you authorize.

Key ideas

01A review: Name the questions it should answer and the material available.
02A change: Name the target and the action you authorize.
03Order: Independent reviews run together; an edit chosen from research waits for the research.

Pick one of these from your own work:

a plan you want questioned
source material you need examined
a project you want to understand before changing it

If the first return misses something, explain what you expected and why. That gives you a concrete repair and better instructions for the next run.

Next comes automation: starting an agreed job at the right time, with a result you know how to inspect.

Quick reference

Where things live

Project agents	.claude/agents/ inside the project
Personal agents	~/.claude/agents/, available across projects
Project Skills	.claude/skills/<name>/SKILL.md
Practice inputs	demonstration-inputs/ in the Section 10 practice pack

Agent file fields

name	The name you use in a prompt, such as customer-plan-critic
description	When the agent is useful; Claude uses it to select the agent
tools	Actions it may use, such as Read or Write
model	Which model runs it; effort can be saved too
maxTurns	Optional limit on rounds of work
skills	Optional list of Skills to preload into the agent

A good handoff

Purpose	What the review or change is for
Starting information	The exact files or material
Allowed actions	Read only, or which file it may write
Result needed back	Format, length and where to save it
Completion	What finished, where it is, what remains open

When something goes wrong

Agent not found	Start a new conversation in the same project and ask again
Wrong agent ran	Name the agent and check its loaded description
Tool missing	Add only the tool the job needs, then retest in a fresh conversation
Unclear return	Say what you expected and why, and compare with the assignment