How to Set Up Your First AI Agent Team (No Code)

0:00Running a single AI agent is cool, but the ones doing real work inside businesses right now are running in teams. I thought putting one of those together would be difficult until I found a tool that does the whole thing with no code. So, in this video, I'll build one from scratch, and by the end, you'll have three agents doing your research, writing, and design without you. The whole team gets built in Hyper Agent, the sponsor of this video, and the link is in the description. Signing in as a Google, Apple, or Microsoft account, or an email address, and there's nothing to install because every agent runs in the cloud and not on your

0:27machine. What opens is home with a box waiting for your first task. And the first thing to get straight is that everything in here runs in threads instead of chats. A thread is where an agent runs a job and everything it produces stays in that thread instead of scrolling away above you. So you can open one 3 weeks later and the documents, the tables, and the pages that agent made are all still sitting in it. That box is also where the first mistake gets made because the instinct is to write one huge prompt covering research, writing, and design and hand the lot to a single agent and nothing in

0:54here stops you doing it. That is exactly what my first build was. The reason it doesn't hold up is that all three sets of instructions sit in the same context window and the model works through every one of them on every run. What comes back is research that reads like a summary, weak source links, and one of

1:08the three sets of rules quietly ignored. I asked it to flag anything a competitor had already covered, and there isn't a single one of those on the page. It costs more as well because that one agent pays for all three sets of instructions every time it runs, where a specialist pays for one. on something that runs every morning, that gap is most of your bill by the end of the month. So before anything gets built, the job I want the team doing is one thing, which is a page waiting on me every Monday morning covering what launched in AI that week with images on it that I can hand straight to an

1:35editor. That job splits into three parts because something has to go and find what launched and back it up with sources. Something has to work out which of it is worth making a video about, and something has to turn all that into the page. So it's three agents: Scout, Ridge, and Atlas. And each one gets one of those parts and is deliberately no use at the other two. Swap those three parts for whatever your week involves because the way a job splits works the same whatever the subject is. So the account is open, the job is split three

2:00ways and there's nothing built yet. Scout gets built first because it decides what the other two ever get to work with. How reliable Scout turns out depends entirely on the job description you write into it. And that is the one part of this worth slowing down for. There's a marketplace with featured agents and skills in it. And I'd come back to that later because building the first one by hand is what shows you the parts. A new agent opens on a row of tabs and the two you set first are identity and model. Identity is the name, the icon, a short description and the system prompt. And Hyper Agents own

2:27docs call that system prompt the job description and that is how to treat it. So Scout's job description says to check a list of AI tool launches and competitor channels every morning, pull the sources, and report what changed on each of them. An agent is only useful if it has a definition of done because Scout has to know when a job is finished without anybody there to tell it. So,

2:44the prompt also says five items maximum. A source link on everyone and no item without a date on it. The other half of a good job description is what to ignore. Scout is told to skip funding rounds, hiring news, and anything without a working product behind it. A research agent with no exclusions on it comes back with everything, and then you are the one doing the filtering, which was the job you were trying to hand over. The model tab is where you choose what runs it. And the list goes from Fable 5 and Opus 5 through GPT 5.6 Soul and Gemini 3.7 Flash all the way down to

3:13open models like Kim K3 and GLM 5.2. Bigger models make better decisions and cost you more. So, this is the one setting worth revisiting once you know what the agent is actually doing. There's also a budget field that caps the dollars an agent can spend per response per run and a separate sub aent model. So, the cheap model does the fetching and the expensive one only handles the hard part. Integrations is a long list of switches, search and the exo research tools at the top, then browser, tables, documents, web pages,

3:38slides, and a full set of media tools. Scout only gets web search and browser, and everything else on that list stays switched off. Every agent also comes with a real browser, so Scout can open a competitor's page and read it itself instead of guessing from what a model was trained on. A research agent that can also publish web pages will eventually publish one without being asked to. Turning on plan mode makes it write up the steps before it does anything, and that's worth having on the first run of a new agent. So, you're correcting the steps instead of the finished work. Scout comes back with the

4:04launches, a source sitting next to everyone, and a line on why each of them matters. That report is useful to me, but I can't send it to anybody else because nothing has turned it into something I can publish. The obvious thing now is to hand scout the writing as well, and doing that is what stops the whole thing working. Every responsibility you add makes the others harder to follow. So instead of adding to scout, I build a second agent called Ridge, whose only input is work scout has already finished. Ridge does not go looking for anything on its own. Its job description says to take Scout's items,

4:31work out which ones are already saturated and which nobody has covered, and come back with the angles worth making, and done is some of those with a reason on each. Hyper agent calls this agent delegation, and one agent hands part of a job to another by mentioning it in the thread, which runs the specialist in its own child thread with its own model, tools, and skills. You can also take the mention out of it by

4:49adding ridge to scouts allowed agents. So the handoff happens on its own during a job that runs while you're asleep. You choose which agents each one is allowed to call and you tell it in its own instructions when handing over is the right move. Two rules decide whether any of this works and I only found both of them after a run went wrong. The child agent only sees the handoff in its own configuration, not the parent conversation. So the brief has to carry the goal, the inputs, and the definition of done on its own. Write it so it makes sense to somebody with no other information in front of them. The depth

5:15is also capped at one level. A specialist returns to whoever called it and cannot pass the job on to a third. So every agent you add answers to the same one. That cap starts to help once you have four agents running because if they could all just grab whatever work came in, you'd never know which one failed. And this way, one agent is responsible for the result. Give them actual names instead of job titles. A team with Scout, Ridge, and Atlas on it is one you can keep track of, and agent

5:36one, agent two, and agent three is not. So instead of one agent doing two jobs badly, Scouts finished research turns up as the material ridge starts from. That's two of the three built. And the last one is what turns the work into something I can send out. The last one is the only one with a design job, and it needs a different tools list to the other two. So, the hyper agent link is the first one in the description. Sign up, and you'll get $100 in bonus [music] credits. Open it now, and you can build

5:59all three of these agents alongside me. The third agent is called Atlas, and the two switches it gets that the others don't are image generation and webpage publishing. What Atlas gets handed is Ridg's finished writeup, and what it sends back is one page a week with a cover image on every item and a headline

6:13that isn't the one Ridge already used. Everything an agent makes goes into the library as an artifact. And those come as documents, tables, web pages, slides, images, and media, maps, files, and hyper apps, which are interactive apps the agent builds when the output needs clicking instead of reading. Watching them appear is how you catch an agent working on the wrong thing. Because you can stop a run 3 minutes in instead of reading a finished piece you were never going to use. I've spent this much of the build on the output format because most briefs never get opened at all. A document gets left in a folder and a

6:42page with images on it actually gets opened and that second version is the one that saves me time in the week. The test I use is whether the output can be sent to somebody with no covering explanation attached to it and a page passes that where a paragraph sitting inside a thread does not. The week's launches come out as a live page with images on it and I can send that link to anyone. The team is built and nothing it learns this week carries into the next one yet. I corrected the same thing week after week and the next run [music] made the same mistake. Learning is where that

7:07gets fixed and it holds two separate things that are very easy to confuse. A memory is context the team should carry. So mine says every item needs a reason it matters. A date and a source. A skill is the method. So it covers how a job gets done, which fields matter, what counts as a risk, and what good output

7:22looks like when somebody reads it. Memories can save themselves because when an agent notices something worth keeping, it proposes one as a draft card you accept, edit, or [music] dismiss. And there's an autosave on the agents knowledge tab you can turn on for your own agents and should leave off for shared ones. Hyper agent can suggest a skill on its own when it spots a reusable method. And you can also ask a thread to propose one off work you've just finished. Either way you edit what it drafts because the first version is always too vague about what it's asking for. Every memory carries an importance rating and anything at four or five gets

7:51recommended to include, so the one instruction that matters never gets dropped on a long run. Skills attach on each agent's own skills tab and they're portable. So the one I write for Ridge goes on to Atlas without me writing it twice and the whole team improves it once instead of you retraining three agents separately. There's a third thing in learning and it's the one I set up last. A rubric is what good looks like written down so work gets scored against a fixed standard. Hyper agent suggests a few rubrics off your work and you hit create on the one you want so it lands in your rubrics tab. From there you can

8:19run an evaluation of any thread against it. Pick the model that does the scoring and it grades the work for you instead of you checking it by eye every time. The shortest gap you can put between two automatic checks is 30 minutes, and there are two ways to run the team that are nothing like each other. There are two ways to make the team run without you, and picking the wrong one gets you a pile of runs with nothing in them. A schedule is a standing prompt on a cadence. So, you name the job, set the time zone, write the prompt, and by

8:43default, every run opens a fresh thread. That standing prompt has to work on its own because there's nobody there at 7:00 in the morning to answer a clarifying question. So, it needs the goal, the inputs, the destination, and what [music] finished means. Live mode is the other one, and you set it up under invocations where it works off a checklist instead of a prompt, waking up on an interval to work through that list and decide whether anything on it needs dealing with. It only comes to you when something on that list needs you. The checklist has to be observable, so a line like check the pricing page for

9:10changes works and make sure everything is fine does not. The floor on that interval is 30 minutes. The default on most new setups is about once a day. And there's a separate heartbeat model for the recurring check, so you aren't paying your best model to find out nothing has changed. The difference between them is the thread because a schedule opens a fresh one by default and live mode stays in the same one so it can compare today against what it already told you. There's a switch in the schedule editor that does something similar, only deliver when something needs attention. And with that on, you only hear from the team in [music] the

9:37weeks there's something to hear about. Those two are part of a set of eight ways to set an agent off and the rest are a thread, Slack, Telegram, email, a web hook, and an MCP call. Web hooks are the one worth knowing about. 30 minutes is far too slow when you want a reaction close to the moment something happens somewhere else. So Scout goes on live mode because noticing things is its entire job. And Ridge and Atlas go on a Monday schedule because their work only starts once there's something new to write about. The team runs itself from Monday morning, and I'm only in it when

10:02something needs a [music] decision. That last part is the one you have to get right because an agent with your Gmail connected can do things you cannot take back. Gmail and Slack both connect in about a minute, and the setting you change afterwards controls what they can do unattended. Integrations get added per agent, so connecting Gmail on Atlas

10:17doesn't hand it to Scout or Rich. Nothing touches your inbox unless you put it on that agent yourself. That second step is there on purpose and it's also the first thing to check when a scheduled run does all the research and then posts nothing. In Slack, you mention the agent in the channel. It goes off and works and it posts the result back and you can have as many agents in a channel as you want. It also runs the other way around because a scheduled job can post its finished report straight into a channel instead of writing it back into a thread nobody opens. So, the report is in the channel

10:43before anybody has opened the app posted by an agent running on its own schedule. The permissions sit on their own tab and they split what an agent is allowed to do in a thread you're watching from what it's allowed to do unattended on a schedule or a live run. The setting to turn on is ask first before external rights. So an agent can research as much as it likes and still has to come to you before it sends, posts, or changes anything in a connected app. Then you go back and switch on just the one single

11:07action you want happening automatically. Every other write stops and asks and that request turns up in command center with the browser tab telling you as well. The dashboard gives you agents, active runs, total [music] runs, and total cost across the top, and it needs attention list underneath with anything waiting on you. That approval step is worth setting up because an agent you have to supervise saves you nothing, and an agent nobody supervises will eventually email a client. That changes again the moment somebody other than you is running it. Sharing an agent through a team gives people run access and not edit access, so an editor or a VA can

11:37use what you built without touching the prompt underneath it. That's why the setup is worth doing properly because you write the job description once and everybody after you just runs it. So, three agents now cover the research, the writing, and the design between them and the page turns up whether I'm working that day or not. And your three agents

11:52will be doing different jobs to mine. The marketplace has featured agents in it already, so your first one can be running before you finished writing a job description. Sign up for a paid plan through the first link in the description and get $100 in bonus credits. Thank you for watching and I'll see you in the next one.