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
Welcome back to Agentic AI Without Coding. We are now in Session 9: Agent Teams, Quality, and Security. In this episode we split one do-it-all agent into a small team of three agents: a researcher, a writer, and a reviewer. Then we connect them in n8n and try them out on Instagram content for Sunset Coffee.
Step 1: One do-it-all agent vs a small team
In Session 8 we built a customer service agent that answers WhatsApp and Telegram twenty-four hours a day. One agent, one long set of instructions, and lots of tools. Problems start when we pile every task onto one agent. It's like telling one employee to do the research, the writing, the checking, and the sending, all at once. The results come fast, but the agent forgets rules, its work is hard to check, and when something goes wrong, we don't know which part to fix. In this episode we learn a tidier way: splitting the work across several small agents. Let's look at a real example. Laras asks one chatbot to do everything in a single request: find ideas, write five captions with prices, match the brand style, check for mistakes, and create visual ideas. The answer arrives fast and looks neat. At the end, the AI even writes: I've checked everything, no mistakes. But look at the first line. The one-liter palm sugar iced coffee is listed at Rp75,000, when the real price is Rp85,000. The second line is worse: free shipping anywhere in Indonesia, a promise Sunset Coffee never made. Laras has to spot those mistakes herself and point them out. But the whole reason for using AI is so that Laras doesn't have to check every word. The AI apologizes and fixes it. But if Laras were busy and didn't read it, that caption could have gone straight online. An AI checking its own writing turns out to be unreliable. Now let's compare the two approaches. Instructions: a do-it-all agent needs very long instructions, and the longer they get, the easier it is to miss a rule. In an agent team, each member only needs short, focused instructions. Checking: a single agent grades its own writing, like a student marking their own exam. In an agent team, a separate reviewer uses a rubric. Finding mistakes: with a single agent, it's hard to tell which part went wrong. In an agent team, you can see each member's result one by one. Fixing: changing one rule in long instructions often shifts other parts too. In an agent team, you only fix the one agent with the problem. The downside is that an agent team calls the AI three times, so it's a little slower and a little more expensive. Picture a small kitchen: someone does the shopping, someone cooks, and someone tastes the food before it's served.
Step 2: Three roles: researcher, writer, reviewer
This is the team we use throughout Session 9. The first member is the researcher. Its only job is to find content ideas from trends and the promo calendar. The second member is the writer. It takes the researcher's ideas and writes captions and visual ideas in the Sunset Coffee brand style: warm, relaxed, and greeting customers warmly. The third member is the reviewer. It doesn't write anything. Its job is to score the writer's work with a rubric: correct prices, the right style, and no false promises. At the end there is still a human: Laras, who approves or asks for changes. Each member receives the previous member's result, does one task, and passes it on. We cover owner approval in full in the next episode. To make the connections clear, define each agent's input and output. The researcher receives this week's dates, the promo calendar, and trend notes. The researcher's output is five ideas, each with a day, a title, a reason, and the featured product. The writer receives those five ideas, plus the official price list and the style guide. Notice that the price list is handed over directly, not left to the AI's memory. The writer's output is five captions, visual ideas, and hashtags. That's all, with no extra comments. The reviewer receives the captions, the same price list, and the scoring rubric. Its output is a score for each caption, notes for improvement, and a pass or revise decision. One agent's output becomes the next agent's input, just like a relay race. Let's write the researcher's instructions. The role section can be short: a content researcher for Sunset Coffee, a coffee shop in Bandung. The data section holds variables the workflow fills in automatically: this week's dates, the promo calendar, and trend notes. The double curly braces mark the parts that change every week. There is only one task: suggest five content ideas, one for each weekday. Notice there's no write, check, or send here. The format is a table: day, idea title, short reason, and featured product. A tidy format makes the result easy for the writer to read. And the most important part, the limits. The researcher may not write captions or mention prices, because those are other members' jobs. An agent with a narrow task drifts off course less often.
Step 3: Narrow instructions for the writer and reviewer
Now the writer's instructions. The role is a caption writer in the Sunset Coffee style: warm, relaxed, greeting customers warmly, and two emojis at most. The data is the researcher's ideas plus the official price list. The task: one caption of sixty words at most, plus one visual idea, for each idea. The limits prevent the mistakes we saw earlier: prices may only come from the price list, and it may not promise promos, discounts, or shipping that aren't in the data. The format per idea is: day, caption, visual idea, and three hashtags. The writer doesn't need to know the trends or grade itself. It just focuses on writing. The reviewer's instructions are the strictest. It may not rewrite captions, only score them. Its data is the writer's captions and the same price list. Its task is to give a score from one to five on four criteria: accurate, on-brand, complete, and safe. We dig deeper into this rubric in Episode 43. Its output is four scores, notes for improvement, and a decision: pass only if every score is at least four. Otherwise, revise. The limit is firm: a single number or promise that isn't in the data drops the accuracy score straight to one. The reviewer is strict on purpose, because its job is to find mistakes. Before connecting anything, check your instructions against these five traits. First, one main task that you can say in one sentence. Second, clear input: what data this agent receives, and where it comes from. Third, a fixed output format, such as a table or a numbered list, so the next agent can read it easily. Fourth, a ban on doing other members' tasks. The writer doesn't score, and the reviewer doesn't rewrite. Fifth, key data such as prices and rules is always provided as data, not left to the AI's memory. If all five traits are met, your team is ready to be connected.
Step 4: Option 1 in n8n: three sequential AI steps
Now let's connect the team in n8n. There are two options. Option one is the simplest: three AI steps that run in order. We start with a run manually trigger, so it's easy to test. The first AI step is the researcher. It holds the researcher instructions we just wrote, plus a choice of AI model. The arrow carries the researcher's output, the five ideas, to the second AI step: the writer. The writer's output, the five captions, goes on to the third step: the reviewer. Finally, the results are saved to Google Sheets so Laras can read them. For now there's no approval step. We add that in Episode 42, and assemble the full version in Episode 45. We click execute. One by one, the nodes get a green check mark. The whole team finishes its work in about thirty-eight seconds. Click any node to see its input and output. This is the strength of the sequential option: if a caption is wrong, you can see whether the idea came wrong from the researcher, or the writing came wrong from the writer. Let's click the writer node to see what's inside. The settings panel opens on the right. The model field holds the AI model of your choice. The system instructions field holds the writer instructions we put together earlier. The incoming message field is yellow because it contains an expression. It means: take the previous node's output, the five ideas from the researcher. You don't need to type it, just drag the output field from the data panel into this box. The reviewer node is set up the same way. Its incoming message takes the captions from the writer. Menu names in n8n may differ slightly from this simulation because the app is updated often, but the pattern is the same. Four tips to make these connections last. First, give nodes clear names: researcher, writer, reviewer. Don't leave them as AI 1 and AI 2, because three months from now you'll be confused yourself. Second, ask for a fixed output format. If the researcher sometimes writes a table and sometimes a paragraph, the writer may misread it. Third, test one node first. Run only the researcher, check the result, then connect the writer. It's easier to find problems one at a time. Once the researcher is stable, add the writer, then the reviewer. Fourth, keep the price list in one place, such as Google Sheets, and use it for both the writer and the reviewer. When a price goes up, you change one cell, and the whole team uses the new price right away.
Step 5: Option 2: an agent calls other agents as tools
Option two is more flexible. The trigger is a chat message from Laras, for example: make three captions for the weekend hampers promo. That message goes to one AI Agent that we call the coordinator. The coordinator doesn't write anything itself. It's like a manager who hands out tasks. Like the agents in Session 7, the coordinator has an AI model as its brain. Its instructions list the team members and the recommended order of work. What's different is the tools. The first tool is the researcher agent. In n8n, an agent can be attached as a tool for another agent, or called through a separate workflow that contains an agent. The second tool is the writer agent, with the same writer instructions as before. The third tool is the reviewer agent, complete with its rubric. And the fourth tool, Sheets, to save the results. The coordinator decides when to call whom. If the reviewer leaves notes, the coordinator can call the writer again to fix it. Each tool agent still uses its own narrow instructions. The only difference is who sets the order. In option one the order is fixed. In option two, the coordinator decides. So which one should you choose? Order of work: in option one it's always the same. In option two, the coordinator decides. Easy to check: option one is very easy, because every step's result is visible on the canvas. In option two, we have to read the coordinator's log to know who was called and when. Revision loops: option two is more flexible, since the coordinator can call the writer again and again. In option one, we have to add revision rounds ourselves as a branch. Cost: option one is easy to predict. Option two can balloon if the coordinator loops through revisions too often. So write a limit into the coordinator's instructions, for example two revisions at most. In short, option one suits routine work like weekly content, and option two suits open-ended requests from the owner. My advice: start with option one.
Now let's look at the test results. The researcher finishes first: five ideas, such as morning coffee for workers on Monday, cold brew for a hot Wednesday afternoon, and hampers on Friday. The writer takes those ideas and writes five captions with visual ideas and hashtags, using the price list pulled from Sheets. Here is the Friday caption. The hampers price of Rp150,000 is correct, the no-sugar version is mentioned, and the free delivery rule is written in full: in town, on orders over Rp200,000. The visual idea is hampers on a wooden table in late-afternoon light. Then the reviewer gets to work and finds one mistake. The Wednesday caption mentions a one-liter cold brew, but the only size on sale is five hundred milliliters. The status is revise. That caption goes back to the writer along with the reviewer's notes, gets corrected to five hundred milliliters, then is checked again and passes. The final result is five captions that passed review with an average score of four point six, all saved in Sheets. The mistake was caught by the agent team, not by Laras in the middle of a busy day. In Google Sheets, the results are laid out neatly by day: the idea, the start of the caption, the visual idea, the score, and the status. Look at the Wednesday row. Its status is pass after one revision. Notes like this help you see which agent makes mistakes often, and which instructions need to be clearer. Five days, five posts, ready for Laras to review. Work that used to take one to two hours of finding ideas and writing now takes just a few minutes of reviewing. If your business isn't a coffee shop, the pattern stays the same. Mr. Hendra at Bimbel Cerdas, a tutoring center, can use the same team for lessons: the researcher picks topics, the writer creates questions, and the reviewer checks the answer key. For your own business, start small. First, just two agents: a writer and a reviewer. You can add the researcher once those two are stable. Second, write narrow instructions with the five traits we covered. Third, save each agent's output, so when something goes wrong you know who caused it. Fourth, limit revision rounds, for example two at most, so costs don't balloon. Fifth, keep a human in the loop before content is posted. The AI reviewer filters out mistakes, but the final decision still belongs to the business owner.
Done!
Let's recap. An agent team splits one big job into narrow tasks: the researcher finds ideas, the writer writes, and the reviewer scores. In n8n, connect them as three sequential AI steps, or as a coordinator that calls other agents as tools. In the next episode we install a safety brake: human in the loop, a way to make the AI wait for your approval before it sends or pays for anything. The agent team we just built will use this brake too. See you in Episode 42.
Practice exercises
Try these on your own after watching to lock in what you learned.
Write narrow instructions for two agents in your business: a writer and a reviewer. Make sure they meet the five traits of narrow instructions.
Set them up as sequential AI steps in n8n, then test them with three content ideas and note the mistakes the reviewer finds.
Add a researcher agent once the writer and reviewer are stable. Save this workflow: we use it again in Episode 45.
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
53 video classes1,279 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.
