At most one email a week. Unfollow anytime with one click.
Transcript
The full narration of this video. The current sentence is highlighted while the video plays; click any sentence to jump there.
Show transcript
Introduction
Hi, and welcome back. We can now plan changes and carry them out step by step. But what if the job is big and parts of it could be done at the same time? In this episode we meet subagents: helpers Claude sends out to work on parts of a task in parallel, which then report back. We'll also create a dedicated subagent of our own.
Step 1: What is a subagent?
Picture the main Claude as a team leader. When it gets a big job, it splits it into several parts that don't depend on each other. Each part is handed to a subagent, a helper that works separately. Several subagents can run at the same time, like three staff members working on different tasks at once. When they're done, the subagents don't send their entire work notes, just a summary of the results. That way the main conversation stays compact instead of filling up with details. The main Claude combines those summaries, and the decision on what to do next stays in your hands. You don't have to tell Claude to use subagents. Claude can decide that on its own for big tasks, but asking for them clearly, as Raka is about to do, makes the results more focused. The first benefit: the main conversation stays focused. Reading dozens of files produces lots of detail, and that detail can stay in the subagent's workspace. Second, it's faster for tasks that can be split. Three checks that each take five minutes can finish almost at the same time. Third, specialized roles. A subagent can be given standing instructions and limited tools, for example read only, with no permission to change files. One important note: every subagent uses its own share of usage. For small tasks, a single Claude is enough. Usage limits and costs can change, so check Anthropic's official pages.
Step 2: Three checks at once
For comparison, Raka once asked for the same check without subagents. A single Claude read all seven pages in a row, one by one, for links, then prices, then layout. The result was still correct, but it took longer, and the main conversation filled up with page content Raka didn't need to read. The conversation's memory space also filled up quickly, making it easier for Claude to lose focus in a long conversation. A check like this is a great fit for subagents: three angles that don't depend on each other, and all Raka needs is the list of findings. The menu and cart feature is finished. Before launch, Raka asks for a thorough check split into three parallel parts, and makes clear that the checkers may only read. Claude explains its plan: three subagents running at the same time, each reading and then sending back a list of findings. The first subagent checks the links and images on seven pages, and finds three problems. The second subagent matches every price and piece of text against the menu data, and finds four problems. The third subagent checks the layout at phone screen width, and finds two problems. All three work at the same time. Everything finishes in about two minutes. The main Claude receives three summaries, combines nine findings, and sorts them by priority.
Step 3: Combining reports and fixing
Here's the combined report. The most important finding: the cold brew price on the promo page says fifty thousand, but the official price is fifty-five thousand. Customers could complain. The second high priority: the send order button is hidden on small-screen phones. Customers can't order, even though that's the whole point of this feature. The Instagram link points to the wrong address, medium priority. The four-megabyte hamper photo makes the page slow, especially on mobile data. And one small typo. The report also records where each finding came from, so Raka knows which checker found it. Nine findings in total, with five shown on screen. Raka decides to fix the two high-priority findings first. The other findings are scheduled for later. The fixes are made by the main Claude, not by the checkers. That way every change happens in one place that Raka is watching. The diff is simple: fifty thousand is replaced with fifty-five thousand. The source of this mistake is interesting: the price was typed directly into the page instead of being taken from the menu data. Claude also fixes the layout so the send button always sticks to the bottom of the phone screen. The commit permission box appears with a clear message. Raka checks on a phone first, then approves.
Step 4: Creating a dedicated subagent
Price mistakes can happen again with every new promo. Raka asks Claude to create a dedicated price-checking subagent that can only read. Claude creates one text file in the project's settings folder. Because it lives inside the project, this file is saved in Git too, and anyone working on this website can use it. Let's read it. There's a name, a description of when this subagent is used, and the list of tools it may use: only reading and searching. There's no tool for changing files. Below that are the standing instructions: only read and report. Now Raka just says its name: use price checker to check the whole website. Twelve prices checked, and all of them match. Claude Code also has a command for viewing and managing subagents. Command names and file locations can change between versions, so check the official Claude Code documentation. From that example, here's the recipe for a good subagent. First, one clear task. A jack-of-all-trades subagent is actually hard to rely on. Second, a description of when to use it. Claude uses this description to decide when to call that subagent. Third, as few tools as possible. A checker only needs to read. The fewer the tools, the smaller the risk of unwanted changes. Fourth, ask for results as a short summary, so reports from several subagents are easy to combine.
Step 5: When to use subagents
When should you use subagents? They're a good fit for checking many pages from several angles, as Raka did. But to change one piece of text or one price, a single Claude is enough. Dewi could use subagents to summarize fifteen branch report files at once, one helper for every few branches. On the other hand, changes whose order depends on each other are better done by one Claude, in sequence. Subagents also suit finding patterns in hundreds of files, for example Mr. Hendra looking for overlapping questions in a question bank. But work that needs your decision at every step is still done together with a single Claude. The screens in this video are simulations, so your screen may look slightly different. Finally, four safety rules. First, subagents follow the same permission rules as the main Claude. So permission boxes still appear, and you still have to read them. Second, never give subagents tools to delete, send messages, or pay. Helpers working behind the scenes should have the least authority possible. Third, check subagent summaries first. As with a report from a new staff member, take one or two findings and verify them yourself before trusting it fully. Fourth, watch your usage. Running many subagents at once uses up your quota faster, so use them only as needed.
Done!
We met subagents, ran three checks at once, combined their reports, fixed the important findings, and created a dedicated read-only price-checking subagent. In the next episode we connect Claude Code to other tools through MCP, such as a browser and the services you use every day. See you there!
Practice exercises
Try these on your own after watching to lock in what you learned.
Ask Claude Code to check your project with three parallel subagents and one combined report
Create one dedicated read-only subagent for a check you do often
Compare the results and time of a normal check with a check that uses subagents
Ask the class material
Type a question or keyword; we find the part of this episode that explains it and jump the video there.
My notes
Notes are saved with the current video time and are only visible to you.
Questions & answers
Stuck on a step? Ask here; the instructor and other learners can answer.
No questions yet.
Your instructor
Cerdas Inovasi Digital Team
Software developers, BIM & 3D engineers, and AI practitioners
52 video classes1,277 episodes
We build software, BIM models, games and AI automations for clients every day, and turn that hands-on work into step-by-step classes: every step is shown on screen and comes with a PDF guide you can keep.
