YT YouTube2Text

YouTube2Text → Transcripts → Claude Code Subagents and Multi Agent Workflows

Claude Code Subagents and Multi Agent Workflows — Transcript

by ExamPro · 8,225 words · 1,267 segments · language en · Watch on YouTube

Open in transcriber Summarize with AI Download SRT/VTT

Full transcript

0:05Sub agents are specialized AI assistants
0:07that handle specific tasks for tasks.
0:09Claude will delegate tasks to agents
0:11based on agents descriptions. Sub aents
0:14run in a single session, can only report
0:16back to the main agent. Uh, and Claude
0:19has multiple ones that are built in.
0:20They're okay, but uh, basically you want
0:22to build your own sub agents. Each sub
0:24agent runs its own context window with
0:26its own custom system prompt. It has a
0:29specific tool access. It has its own
0:32independent permissions. So let's look
0:34at some of the built-in sub aents. We
0:36have explore. It's a fast read only
0:38agent optimized for searching, analyzing
0:39code bases. We have plan. It is a
0:42research agent used during plan mode to
0:45gather context before presenting a plan.
0:47In fact, I believe that we have been
0:49using this without you being aware of
0:51it. We have general purpose one. So,
0:53it's a capable agent for complex
0:55multi-step tasks that require both
0:57exploration and action. Then you have a
1:00bash for running terminal commands in a
1:02separate context. Status line setup for
1:05when you run the status line and the
1:07claude code guide. So when you ask
1:09questions about cloud code features, um
1:12we can create custom sub aents. So that
1:15will be in our agents directory. They
1:17are markdown files. Everything's
1:18markdown files with front matter
1:20information. Uh you can see what tools
1:22it has its model description but
1:24actually has a lot of functionality and
1:26so here's like a full one where we have
1:28a senior code reviewer. So we obviously
1:30have name description which will trigger
1:32it disallowed tools the permission mode
1:35uh so like whether it's don't ask max
1:37turn skills that it can use which is
1:39over here uh MPC servers which it has
1:42access to hooks right memory so is it
1:47project based uh whether it's allowed to
1:50run the background or not or and its
1:52level of isolation like does it work in
1:53a separate git tree so very very
1:56powerful uh rolling up a lot of these
1:59things together. Uh we have our sub
2:02agent. Okay.
2:08All right. So let's take a look at uh
2:10getting some sub aents running. So sub
2:11aents are interesting because they are
2:14in their own basically isolated
2:16environment uh having their own uh
2:19context window. Um and obviously like if
2:22you wanted to run multiples doing
2:24different things that would be really
2:25really useful.
2:27um the they're not going to coordinate
2:29with each other. That's going to be a
2:30team of agents. But um if there's
2:32individualized tasks that they need to
2:33do, they're going to really excel at
2:35that. So, you know, building an agent is
2:38quite a bit of work. So, maybe we could
2:39try to find something that already
2:41exists. I just went to the GitHub, typed
2:42in anything. There's something here that
2:43says awesome uh agents with 100
2:46specialized code agents for sub agents.
2:48It's by Volulta agent. I guess they're
2:50trying to get you to use their uh
2:52product, whatever that is. um which is
2:55over here and looks like something. I
2:57don't know what it is, but it's like
2:59whatever this is. Um which is cool,
3:02whatever that is. But I'm only
3:04interested in these agents. So, um let's
3:08go ahead and see if we can get this
3:09installed. And actually, I guess it's a
3:11plugin. So, we're covering plugins
3:13before we do the plug-in section, which
3:16is fine because, you know, the less I
3:18have to do for labs, the better. as this
3:20course is getting long long long in the
3:23tooth here. Okay, so I'm gonna grab this
3:25and we're using basically the Claude
3:27plugin uh marketplace and adding a volt
3:30agent which will bring them in here and
3:32so now we have that in there. Um now it
3:35says Claude plugin install plugin name.
3:37What are you talking about? We just did
3:38it.
3:40Um so oh over here. Okay, so this is
3:42where we can add our specialized uh
3:45agents. So if we go down below here we
3:48got a bunch. Okay, we got language
3:50specialists, core development, um
3:52infrastructure, quality security. Wow,
3:55look at all these data AI questions. Are
3:58they any good? I don't know, but um I
4:02guess we'll find out. So,
4:08what can we do? Um,
4:12let's go ahead and install this one,
4:17which is going to be
4:20Claude plugin. Yeah, claude plugin.
4:24Install
4:25this one. Core dev.
4:28Okay. And then we'll go over to this one
4:31here.
4:32Language specialist. That's kind of
4:34cool.
4:36Claude plugin. Install this.
4:41Right.
4:44So now we have a bunch of agents that we
4:49can use. Um, do we have a way to see
4:52where these agents are? Let's take a
4:54look here. Agents.
4:56Here we go. And so now we can see all
4:57the agents that are available to us.
4:59They're all here, right? So what I want
5:02to do is I want to invoke one of these
5:04agents. And um
5:11we have all these experts here.
5:14So I'd like to ask the the the Golang
5:18Pro expert if it can review my code. So
5:21we'll say uh can the um Golang Pro
5:27expert agent
5:29review my Golang code backend? Okay. And
5:33I want to see if it will actually
5:34trigger
5:36that
5:38um sub agent.
5:42And look, it's triggered it. Okay. And
5:44now it's working in the background. Can
5:45I go check tasks? Is there anything in
5:47task? No, no, no. I was hoping there
5:49might be something there. But you can
5:51see it is now off the races. And I don't
5:53think I'm held up here. I think I can
5:55continue on if I want to.
5:57Or maybe I can't. Maybe we can't. But I
6:00might go over here and we'll just say
6:02can you also
6:04Oh, is this one queued up?
6:07Oh,
6:09I maybe because this is not running in
6:10the background. Just a moment here. All
6:13right. So, it's coming back with its
6:14findings here. And so, here's here's the
6:16full review uh from the Go Expert. And
6:18so, here it's suggesting these things.
6:21Um
6:24is not a prompt based skill. So, I don't
6:26know. So it seems like there's some
6:27skill that it has, but no task currently
6:29exists. All right.
6:32And so here we've invoked that agent.
6:34One second. All right. So now that we've
6:37done that, um what I would like to do is
6:40get more than one agent uh running at
6:42the same time. Okay. So the thing was is
6:45that um it did hung up hang up the loop,
6:47but I want to see if we can do more than
6:49one. So we'll go over to here back to
6:50agents. And we clearly have a bunch,
6:54right? And so what I want to do is I
6:57want to get two different languages
7:00and I want them to report back to me
7:02because we have a Rails expert and then
7:04we have a uh PHP expert. Okay. So, I'm
7:08going to go here and say, uh, can you
7:11let the Rails expert agent
7:15and the PHP expert agent look at the
7:20Golang
7:22backend and suggest
7:26uh if it would be easy for them to port
7:31it over to their specific language. And
7:34so, what I'm trying to do is trigger
7:35both of those separately. I'm not trying
7:37to start a team as that is a separate
7:39thing. It's a it's a newer feature agent
7:41teams. So it says I'll run them in
7:44parallel to review them. Okay. So that's
7:45how we're going to get them to run. So
7:47now they're both running. I keep
7:49thinking if we go to tasks, we'll be
7:50able to see the running tasks, but I
7:53guess Oh, nope. There it is. Okay. Yeah.
7:54So there are there are the tasks and
7:57there we can see them. Okay. So that's
7:58what I wasn't sure about. So here it's
8:00running those two agents and we are just
8:02chilling out waiting for those to go.
8:04both ages are running in parallel uh in
8:06the back end. And so if I go back here
8:08and check the tasks again, we can see
8:11them. We can even click into them and
8:14it's showing us our progress in terms of
8:16what they're doing. So that is pretty
8:19pretty cool. Um I wonder if we can get
8:22other information from tasks. Tasks.
8:25Nope, that's about it.
8:29So, we just want to go here and check or
8:32you know, I'm just going to stop and
8:33wait until they come back. Oh, okay. The
8:35Rails one came back. So, Rails expert
8:37reviews the uh So, we have go backend
8:40for reporting completed.
8:43All right.
8:47So, easy. The backend is so simple.
8:49There's almost nothing to translate,
8:51just scaffolding. And so, it has an
8:52example. Docker changes amount of
8:54changes. And then for PHP, they came
8:56back and they said, "Easy." And then it
8:58does that. So, isn't that really cool
9:00that we could quickly port something to
9:02multiple languages extremely quickly?
9:04And then now we get a summary. Oh, this
9:06is awesome. Look at this. So, we have
9:08it. Easy, easy, very low. Uh, whatever
9:12moderate, that's not very clear. PHP
9:14definitely uses less less than Rails.
9:16Um, and then how it would end up
9:18implementing and why it would do it.
9:20That is a cool example of it. So yeah,
9:23we found a sub aents that we can utilize
9:25and install. We could create our own
9:27custom agent. It's not like it's that
9:29hard. We've already created other stuff.
9:30I'm sure you have an idea of how we can
9:32do it. I just don't have a use case for
9:34one right now. Um and I really just
9:36wanted to show you sub aent use and how
9:39we could work with it. U but understand
9:41that like that's going to consume more
9:42of our credits a lot quicker. I'm not
9:44sure we we keep track of that within our
9:46single SE session here. So, if we go to
9:49context, because this is going to show.
9:52Oh, that didn't help at all. Context
9:54again. Context. And maybe because it's
9:56so darn long. If we go here,
10:00we can see custom agents,
10:02right? And maybe when it's actively
10:04running.
10:06Oh, because we have so many available to
10:07us now. And so here we can see across
10:10the board maybe it's agent what it could
10:12actually use.
10:14Oh, did it actually load them all in to
10:16be able to do that? I'm not really sure.
10:18Um, and so we might go here and just
10:21check our usage.
10:23I'm still doing pretty good. But anyway,
10:26that sub agents, that was fun.
10:34Let's take a look at the feature called
10:36team of agents or agent teams. This lets
10:38you coordinate multiple cloud code
10:40instances working together. So, one
10:42session acts as a team lead. Uh,
10:44teammates work independently. Agents
10:46have their own isolated context but
10:47communicate with each other. You can
10:49interact with any teammate without
10:50interacting with the lead. Uh when to
10:53use team agents? Well, when you want to
10:54investigate different challenges to a
10:56problem at the same time independently
10:57work on different features that don't
10:58overlap. Work on the same problem but on
11:01different theories to see the outcome.
11:03Working on isolated layers like front
11:05and back and tests. And so sub aents is
11:08when you need uh you need quick focus
11:11workers that report back. So here you
11:13can see they are all doing their own
11:15stuff and then reporting back to the
11:16main agent. And then an a team of agents
11:19is when the teammates need to share
11:21findings challenges with each other and
11:22coordinate with their own and they have
11:24the shared task of a task list of things
11:28they need to complete. Notice that they
11:30can talk to each other, right? So that's
11:32the key thing here. They're talking to
11:34each other. Here they're not talking to
11:35each other. They have a shared task
11:36list. Each of them have their own task
11:38list. they're not going to delegate to
11:41each other. Um that so agents teams are
11:44disabled by default. So we got to turn
11:47them on and there's two different ways
11:48that we can do that. It's just an
11:49environment variable. Um to create an
11:52agent team simply prompt it, right?
11:54Saying I need create an agent team to
11:57explore this from different angles,
11:58right? And technically we kind of did
12:00that with the sub agent, but maybe we
12:02experienced something a little bit
12:03different because they could talk to
12:04each other. An agent team consists of
12:06team leads, a teammate, a task list, and
12:08a mailbox. Uh teams and tasks are stored
12:11locally. So we can actually find them
12:14over here. And the task list is stored
12:17over here. And teammates start with a
12:19lead's permission settings. If the lead
12:22runs with dangerously skip permissions,
12:24all teammates do so. So it can be super
12:27dangerous
12:28um in that case. But you know, it
12:31depends on what you're doing. There you
12:33go.
12:39Okay, let's go ahead and get team of
12:42agents working. So, we're going to have
12:43to enable it. Um, and so we need to set
12:46that flag there. I'm going to go to
12:48Claude. Oops. Come on, Claude. Wake up.
12:52Let's close some tabs here so we have
12:53less of an issue. Let's go ahead and say
12:55new terminal
12:57and try this again. There we go. go and
13:00say, can you enable uh team agents for
13:04me in my settings file?
13:07Okay, so I need that wherever it needs
13:09to be enabled, we need to do that
13:12and
13:15it's going to get that enabled.
13:23And so we'll say yes,
13:26it's adding that where I'm hoping over
13:28here so I can see it, but it's okay if
13:30it's not there.
13:36And it's looking up the documentation to
13:37figure it out. What's interesting is
13:40like if you notice there's like a lot of
13:41MDs that it's fetching, which is really
13:43really useful.
13:49Okay. And so there it is. All right. So
13:52that's what we're looking for. It is now
13:54enabled or it should be. And so the next
13:57thing we need to figure out is what is
14:00it that we're going to run here. It's
14:02still computing something. Okay, so it's
14:04added it. That's great. Um, and so we
14:07are uh we have these uh Volt agents.
14:11It's like Volt agents uh GitHub. I'm
14:14just looking up the page again because
14:16they do work. They're pretty
14:18straightforward.
14:20Okay. Um,
14:23and
14:26I'm like just kind of wondering, you
14:27know, what's something we can coordinate
14:28with a bunch of these
14:32all the stats? I don't want that. I just
14:33want the Sorry, folks. I just want your
14:35agents. I just want your agents. I You
14:37probably have a good product, but you
14:39know, I don't know. I I don't have time
14:40for that. So, what we're looking for is
14:44um enough things to go on here. So we
14:48have API designer, backend developer,
14:51front-end developer. I wish there was
14:52like a product designer because we could
14:55be like, "Hey, can you go design this
14:57product?" And then we can delegate to
15:00other people.
15:02uh
15:04language specialists,
15:06infrastructure specialists,
15:08data and I developer experience,
15:12build a system specialist, um
15:16specialized domains. So what I'm going
15:18to do here, I'm going to go over here
15:20and I'm going to ask it
15:23uh we'll make a new repo. This is what
15:24we're going to do. We're going to go
15:25back a step and we're going to say uh
15:28make diir link tree uh clone 2.
15:32[laughter]
15:33Let's make another link tree. Why not?
15:36And I I called it link tree. So, let's
15:38try that again. Uh link
15:41tree auto link tree
15:45clone super clones. Super clone. Why
15:47not? Yeah, let's do let's call it that.
15:50And what I'm going to do is I'm going to
15:52open this up.
15:55I also want to know like is that setting
15:57only in this one? No, it's it's
15:59everywhere. So, we should have the
16:01functionality we need. So, what I'm
16:03going to do
16:05is
16:08uh open up terminal here
16:12and
16:16going to make a new read me here.
16:21Okay. I'm going to say uh because I have
16:23a task. I'm like this website has a has
16:28many
16:31uh sub aents
16:35we can use to accomplish
16:38a job.
16:40What I want to do,
16:42I want to build
16:46a link tree clone
16:49using agents using these agents
16:54uh with
16:57team agents mode
17:00in cloud code.
17:06So
17:07I need you
17:09to make a plan for uh assembling
17:14a uh team of agents.
17:18So here, so yeah, I got to write this a
17:21little bit better. Okay, so goal
17:26build a link tree clone. A good looking
17:31link tree clone.
17:35Um
17:39technical goal.
17:42What was the business goal?
17:49Build
17:51build the product build the app using
17:53sub aents
17:56using a team of agents
18:00uh from this list from this repo.
18:07Okay.
18:08So
18:11tasks
18:13uh
18:15create a plan.
18:17So look at the possible agents
18:23and assemble a team and then create a
18:29plan to help them execute
18:33uh the goal. Okay. And we'll say
18:36technical requirements.
18:45This is a simple MVP.
18:48Um, please use Golang for the back end.
18:53Use Docker in order to And by the way,
18:56if we're going to use Docker, I'm going
18:57to make sure I don't have any containers
18:58running right now. Uh, I'm just going to
19:01kill them all here.
19:03Stop.
19:07This is a simple uh MVP.
19:11Please use Golang for the back end. Uh
19:14please use
19:17uh plain JavaScript
19:21well I guess you use whatever wants use
19:23whatever you want for the front end
19:33um
19:34it should all run with docker compose
19:41you need to
19:43test the work.
19:47Okay, so here is the readme. And so
19:51we're going to go here.
19:54And this this extension is not very
19:55useful to me anymore. Uh okay, so we'll
19:58go claude
20:02and
20:03um I don't know if we're going to end up
20:05using all of our stuff doing this. I
20:07don't care if we do. I just want to see
20:08if it'll execute. Um, and so we're going
20:11to go ahead and say um
20:15I need I want so please
20:20read the read me.
20:23Okay, I need a team assembled.
20:31And so it's going to go off and do that.
20:35I probably should have told it to go in
20:36plan mode because it might actually try
20:37to do this first.
20:39Yeah, I'm gonna stop that for a second.
20:43I'm gonna tell it to plan first. I'm
20:45gonna stop it. I'm gonna stop it because
20:47that could get that could be pretty bad
20:48if I if I don't do that.
20:54Please read the read me.
20:58Uh, and
21:00okay. And so that's all it should really
21:02need to know.
21:06And let's see what it does. The only
21:08thing is I haven't installed all the
21:09plugins, right? So, there are some
21:11plugins we don't have. We'll say yes.
21:19The only thing I can see here is me
21:20hitting like yes. Yes. Yes. Constantly.
21:24Do you want to allow Claude to fetch the
21:26content?
21:28Yes.
21:30So, now it's going to that website. It's
21:31going to see all the possible agents are
21:33there.
21:36But I think this is kind of fun. It's
21:38just like go hire go hire the sub agents
21:40and then go have them do that.
21:47But yeah, this this has me thinking
21:48because like you know I have my own
21:50custom JavaScript framework. I we have
21:52very specific infrastructure. So I could
21:54see you know people spending their time
21:56just making their sub agents or their
21:58agents that could be execute on this
21:59stuff. And so we're just saying yes and
22:01letting it figure some stuff out. I'm
22:03still waiting for a plan. Did it come
22:04back with a plan? I gota I gotta pay
22:07close attention here as this is going.
22:09So I'm going to open this up.
22:12Is it already writing code? Like why is
22:14it going to link tree?
22:18Uh
22:23why why? Okay, sure. Well, why is it
22:25looking at that?
22:28Hold on. Cancel.
22:31Where is this? Where's this running? Is
22:32this running in the right directory? It
22:34is. Okay.
22:36So we'll say plan
22:39read this file. So read the readme.
22:43You need to make a a plan for a team of
22:48agents. Okay. You are building from
22:52scratch
22:54uh in this uh folder. So you don't need
22:58to look at any other code.
23:02Okay. So we'll go ahead and try that
23:03again.
23:06because I don't know why it was going
23:07outside of the current current
23:09directory. It didn't need to do that.
23:13Mhm.
23:17So, we'll say yes.
23:22Yes.
23:38Mhm.
23:42And we are just chilling out here
23:44waiting to see what happens.
23:48Okay, so now it's going to write us a
23:49plan. I I do like executing that in the
23:52other menu because then it has a
23:53markdown file. I probably could have
23:55said like make a plan in the markdown
23:56file and then have it. Um that probably
23:59would have been smart to tell it to do.
24:01I didn't think about doing that.
24:04Um,
24:08but we'll have to wait here. The only
24:10challenge is like it's hard to see what
24:12what's happening with the context. So,
24:14we have a plan. So, we'll go up to here
24:16and we'll take a look at the plan. So,
24:19build an MVP link uh link tree clone.
24:21The app needs a Go API backend modern
24:23frontend docker compose. The project
24:25currently contains only readme. And here
24:27we have aid a uh API designer, Golang,
24:31react specialist, docker expert, QA
24:34expert. Um
24:37the outline looks good. The endpoint
24:40looks good. There doesn't seem to be any
24:42like designer.
24:45So it says clause written the plan. Yes.
24:47And clear. Um type here to tell what to
24:50change. I I are there are there any
24:55design agents? Uh I noticed everything
24:59noticed the agents are all agents chosen
25:04are all um
25:07what would you say all
25:10uh tech uh tech like implementation
25:14focused?
25:16Um,
25:18also can you output the plan to a
25:22markdown file
25:26called plan so I have it for later in
25:30this repo?
25:46generated it. Okay. So, we'll go ahead
25:48and do that because I think there might
25:49be other ones,
25:51right?
25:58And so that's
26:00what we're doing. Um,
26:06there's even a content marketer
26:07specialist, a product specialist.
26:09There's all sorts of fun stuff here.
26:11Okay, so here it says build an MVP link
26:13uh uh link tree clone from scratch. The
26:15app needs to go to a backend API modern
26:18application. And so here is the new
26:21file. Do we even have it? I guess it's
26:22trying to save it.
26:24Yeah. Oh, yeah. We'll let it add it.
26:25There we go. Sorry, I'm just trying to
26:27read it there. And that's actually not
26:28its review.
26:30And so here we now have a UI designer,
26:33but like there's still other stuff
26:34that'd be really interesting to check
26:35out. But right now, I guess this doesn't
26:38really matter. Um, so
26:41yes, clear context. I think it's clear
26:43enough.
26:46I don't want to auto accept edits.
26:54Why would you like to proceed?
26:57Yes. Manually approve the edits. Um
27:05I guess so. But we need the team of
27:07agents set up. So
27:10I'm going to hit escape here because
27:11it's fine. Like we now have a plan. And
27:14the plan is fine. I like the plan. Um,
27:18and so what I'm going to do is I just
27:20need to make sure that all the agents
27:21are installed. Um,
27:24so
27:26the plan mentions
27:30I should be able to just look this up
27:32really quickly. So we have here
27:34who's in our team. We have
27:38um
27:40API design.
27:43So I just want to make sure we have all
27:44these
27:48I guess this is like core. These are the
27:50plugins right there. Actually, we can
27:51see them right there. Um, so
27:56what is it? Claude login. What? What is
27:58it? Um,
28:02also kind of wonder if we're able to
28:04like install the sub one there. But we
28:06have cloud plugin
28:10install. I'm just going to make sure
28:11they're all here. So, we already have
28:13the core. I think we already have lang.
28:16We're going to need infrastructure. So,
28:17I'm going to go ahead and install that
28:18one. I'm going to bring in this one
28:21next.
28:24Okay.
28:26And I'm pretty sure we already have dev,
28:28but I'm going to grab it just in case
28:30because I don't trust it.
28:35And I we already have we already have
28:37Golang, but I'm going to do it again
28:38anyway just so that um we 100% have
28:41them. Okay.
28:44So, what I'm going to do is go back to
28:46here
28:48and
28:50I need you to execute
28:55the plan
28:59uh with a team of agents.
29:02The team of agents
29:05uh required are in the
29:09document
29:12and we have a plugin installed.
29:16So they should all be available.
29:30Okay.
29:34All right. And so hopefully that will
29:36just work.
29:38Um but what should be interesting is the
29:40fact that there should be things that we
29:41can observe. I don't know if we have
29:42like have access to mailbox here. We
29:44don't. Um but maybe we'll be able to see
29:47the tasks because I'm not sure exactly
29:48how we're going to monitor them, right?
29:51Maybe that's what that vault product is.
29:53I don't know. [laughter]
29:55But uh you know trying to get anyone to
29:56sponsor anything is like nearly
29:58impossible. So, um, so we have home,
30:00Andrew sites. So, we'll say yes, always
30:03for this project.
30:07The only thing I don't know is like how
30:09are we going to just get it to not
30:10prompt me so much? And that's kind of a
30:12trick because you don't want to have
30:13dangerously uh running permissions.
30:16But anyway, it's off to the races. Let's
30:18go ahead and check out tasks. Are are
30:19there any tasks queued up? I don't see
30:21anything.
30:23Um,
30:25is there a repo for the team? I don't
30:27see it.
30:34So, the only thing I don't really know
30:36is like how are we going to monitor
30:38them?
30:40So, we have control your team here. Talk
30:42to teammates directly. Uh-huh. Well,
30:44how?
30:46Um because they have like split windows
30:48and stuff like that.
30:54Well, here's a Well, it must be. So it
30:56says launching the design in parallel.
30:58So
31:00I mean is this running in teams? Hold on
31:02here.
31:11Is is it running in teams? I can't tell.
31:15So after you uh after enabling your
31:17agent, tell it to create a team. From
31:20there it will create a team and a shared
31:21task list. That's what I'm looking for
31:23right now.
31:24Crl T.
31:28So where is it?
31:31Uhhuh.
31:34The default mode is auto, which is used
31:36to split PES if you're already running
31:38inside a T-Max thing. The T-Max settings
31:40will enable the split mode. Um to force
31:43claw pro uh force in process mode, do
31:46this.
31:49So hold on. I'm going to stop here for a
31:52second.
31:54Hold on. Hold on. Hold on. Hold on. Hold
31:56on. Hold on. Hold on. Hold on.
32:01All right.
32:03I'm trying to get back here and I can't.
32:05There we go. Okay. Debug. Are you Were
32:10you running
32:13in team agents mode?
32:16It seemed like you were just running
32:23sub agents, right? Because we should be
32:25able to access that task list somewhere.
32:30And so I'm going to ask it like, hey,
32:31are you actually doing what I'm asking
32:32you to do?
32:35Yes, I was using the agent tool with the
32:36sub agent type to launch sub agents, but
32:39looking at your available deferred
32:40tools, you have team create, team
32:42delete, which is in team modes. That's
32:43the correct approach here. Let me check
32:45your settings to confirm.
32:48So it is enabled. I should be creating
32:50teams and sending messages instead of
32:52agent tool. So, let me redo phase one
32:54for the teams. And so, it didn't do it.
32:57Isn't that interesting? Okay. So, now we
32:59got something completely different.
33:00Look, it's actually building out a list.
33:02So, there we go. That's what I wanted to
33:04know about like where was that thing?
33:06They do say there's a hotkey in here for
33:08control T. Um, I don't really want to
33:10fiddle with the UI here as I feel like I
33:12could easily break it. um
33:16wherever it was here,
33:21but it's supposed to have like split
33:23screen and because we're not using
33:24T-Mox, I feel like we are not getting
33:27the full uh full power of this because
33:29we're running in Visual Studio Code. So
33:31maybe this should have been something we
33:33launched in uh terminal and then made
33:35sure that T-Mox is installed. I mean
33:37technically T-Mox could be installed on
33:39Visual Studio or sorry within WSL 2. So
33:42maybe that's a non-issue. Okay, but we
33:44go here and it says here now creating a
33:48task list. All six tasks created now
33:51spanning phase one. Phase one is running
33:53both AI designer in the live tree. Once
33:56they report back, we'll move on to that
33:58one. So, while that's happening, I'm
33:59going do control T. Uh I thought it was
34:02control T. Let's go back over to here
34:05because it is running in the background.
34:06Oh, it's coming back already.
34:09And now it's sitting in the visual
34:10design, but it's moving so darn fast, we
34:13can't even see what's going on. Um,
34:20where is this task list? Task list.
34:23Task list. Task list. Here. Control T.
34:27Well, I was pressing that, but I wasn't
34:28seeing anything.
34:32So, look. Control T. Oh, there it is.
34:34Okay, great.
34:38Oh, it even tells us which one's doing
34:39it. That's super cool.
34:43Uh, the only thing is like how much of
34:45the context are we using? Can I Can I
34:47run that while we're doing this context?
34:51No, we can't really. So, I'll just hit
34:53escape and get that out of Q.
34:59We'll do control T again. I don't know.
35:01We just chill out, I guess.
35:08So, I'm not sure who's approving stuff.
35:10It's not like I gave it dangerously
35:11skip.
35:15Also, like at the end of it, is it going
35:17to report to us um how it turned out?
35:23Okay, so phase one done. Now, it's
35:25running phase two. So, all three agents
35:27live in parallel. I while we're looking
35:29at that, let's go take a look at our at
35:30our phases here.
35:33So run all three agents in parallel. So
35:35do the back end here and the phase three
35:37is testing. And so that's our execution
35:40order. What's really interesting is like
35:41when we did the boot camp last year, we
35:43kind of were replicating this but we're
35:45doing a lot of manual work.
35:49So now it's just waiting on some
35:50members. I wonder if any of them are
35:52stuck. So here it's asking for
35:53permission. We'll go ahead and say yes.
35:56We'll also give that permission for the
35:57front end. So it says yes.
36:00And this thing's kind of stuck right
36:02now, but I don't want to touch anything.
36:04So,
36:06I'm doing control T and I just can't get
36:08it to show me the um anything right now.
36:16There we go. So, we'll say yes.
36:22Okay, now we're waiting on that.
36:28But yeah, it'd be really cool to have
36:30like all the split windows and they're
36:31going and then we're just like I don't
36:33know. We feel like we're in the Matrix.
36:35[laughter]
36:37Uh
36:46Mhm. There it goes. Okay.
36:52And while it's running, I'm gonna go to
36:54Claude IO because like I feel like how
36:56else am I going to check my usage as as
36:58this is going. Um, so if I go into
37:02settings here, I'm just doing this off
37:03screen and we go to usage.
37:06Here we have our usage right here.
37:08Right.
37:11So it's going
37:14and now it's running phase three. So now
37:15it's in the testing phase. I'm not sure
37:17if this like updates in real time. So,
37:19like if I keep refreshing this, will it
37:20show me?
37:24Doesn't seem to be going up.
37:29I just wish I had better visibility
37:31here. Yeah, as this has timed out,
37:41I think like having some kind of
37:42interface to watch it would be really
37:43useful.
37:45Now, I'm just really curious what that
37:46volt agent thing is.
37:52Does this display build enterprise
37:53multi- aents development observability
37:55deployment one platform? So I I swear to
37:58you they're not a sponsor. They're not
38:00anything. I'm just It's just because we
38:01saw their thing.
38:04Uh yeah, I don't know what this is. I
38:06can't tell. It doesn't tell me anything.
38:10Like it's a nice looking site, but it
38:12doesn't tell me anything. Coming back
38:14over here, we'll say yes. I should have
38:16put those notifications. So, I get those
38:18little uh dings in my ears because I
38:19can't even tell when it's asking me to
38:21do stuff.
38:25Okay.
38:35Um and again, maybe we go back and check
38:38our usage here. Give this a refresh
38:40here.
38:42So, now we're at 38%. So, it is
38:44definitely changing the value. Oh, and
38:46then we had to do that.
38:55Mhm.
39:21All right.
39:23Uh, there might some structural issues
39:24there. Structured message cannot be
39:26broadcast. I'm not sure what. It's
39:28running. Oh, it's not running. Is it
39:30running?
39:32I just have a hard time believing that
39:34it's already running. Okay, so we do
39:35have well we have one container running
39:38over here and all six teammates have
39:41have shut down the it's built and there
39:44the project is ready is here you can run
39:46that um
39:50so it's telling us to run it. So I let's
39:52go here and say let's hit escape
39:56and we'll say okay great can you
40:01go can you run run the app then
40:06unless it's already running I can't tell
40:08because we do have this one here but it
40:10doesn't have a name.
40:14So that's probably what it is. It's
40:15probably already running right there.
40:37All right. So, it's building it. I I
40:39mean, I would use a Docker Compose to do
40:40it. It's a little bit different how it's
40:41working. We'll go ahead and say yes.
40:45I could have just checked the URL that
40:46entire time. I don't know why I didn't
40:47do that.
41:01By the way, our usage is only at 39%.
41:03So, it's still pretty good for building
41:05a full thingy.
41:10I'm not sure why it's having such a hard
41:12time running.
41:14I guess it's because it's building it. I
41:15could have ran that manually.
41:22Okay. Okay. And so now it should be up.
41:25Uh it says it failed.
41:29The app is now running. Okay. So we'll
41:31go ahead and go to local host 8080.
41:34Um or sorry 3000.
41:383000 here.
41:41I'm just doing this off screen.
41:46There we go. And they completely built
41:48it. Um, is it as nice looking
41:52as the other one that we saw because
41:54remember there was the um the sub agent
41:56that we saw from um what you may have it
42:00but maybe there's a reason for it and so
42:02probably what I would do to get like a
42:05full understanding of it I would
42:06probably ask it to report back
42:08information. Oh, we do have a spec here.
42:09That's what I was looking for. So here
42:12we have this um and what it uses its
42:15specification why it uses it and so like
42:18as a client I would like to know its its
42:22direction and probably I would like it
42:24to like explore different kinds of
42:26designs and come back to me and then
42:27approve it right so I can already kind
42:30of see a process on how to make this a
42:33little bit better but um yeah it's cool
42:37so that's a really good um uh a good
42:41case. I again I still don't think we
42:42have full visibility, but the boot camp
42:45would be something where we would go and
42:47really make this something, you know,
42:51that you would actually do. Okay, we're
42:53just playing around here. Uh this
42:55obviously wasn't hard for me to figure
42:56out. Um but I would want to take, you
42:59know, my knowledge of doing this stuff
43:01like running a firm or working with a
43:03team and stuff like that and really
43:05looking how do we uh make this for real.
43:07Okay, but anyway, this is a great way to
43:10wrap up this course. So, I hope you had
43:12a lot of fun and uh this was really
43:15exciting for you this last part. Okay,
43:17chowo chow.
43:23So, context fork allows you to tell a
43:27skill to uh run as a sub agent. And so,
43:31the advantage of that is it's not going
43:32to hold up your loop. So, you can
43:34continue on with other work as it is
43:35doing that on the side. And so you can
43:37just set that in the front matter. It's
43:39as straightforward as you imagine. Uh it
43:42probably make more sense when we start
43:43working with sub agents, but I need to
43:44mention this because it is a very useful
43:46feature and we will we'll we will give
43:48it a go.
43:53Hey folks, it's Andrew. In this video, I
43:54want to figure out a strategy that we
43:56can use to back up our uh conversations.
43:59There isn't a um a default skill or or
44:03functionality that can do that, but I
44:04think this would be a great opportunity
44:06for us to go build out a skill. And um I
44:10think it's something that uh you know,
44:12you might want to use. So what I'm going
44:13to do is I mean we can do this in any
44:16repo to be honest because we just need a
44:18generic skill for this, but we need the
44:19skill creator um uh skill creator uh
44:23task. What I'm going to do is I'm going
44:24to go over to anthropic uh skills. It's
44:27a repo over here. And so what I need, I
44:30don't want all these skulls. I just need
44:31the ones that I need. And the one that I
44:32need is specifically
44:35skill creator. Um, and what's
44:36interesting is if you actually use
44:37Claude, Claude will actually like
44:39cloud.io, it will actually use this one
44:41to help you build your skill creator or
44:43your skill. Um, so I'm going to go into
44:46CD
44:47into our cloud directory here. I'm just
44:50going to open it up. So I'm going to
44:51type in code period. And this is not
44:53really where I want to work, but in here
44:55we should have skills. I want to see if
44:56I already have that in here. So, I'm
44:58looking for skills. Looking for skills.
45:00Have I not done that here? I think it's
45:02because I've been building so many
45:03skills on my uh [laughter]
45:06other computer. I just assume that
45:07they're in the course here. Uh, by the
45:09way, I have my little space heater
45:10running, so you might hear it. It's very
45:11cold in my office. I apologize if it is
45:13distracting. Um, if it is, tell me and
45:16I'll promise not to do it ever again.
45:18Um, but anyway, so we it's going to go
45:20in a folder called skills. I need to get
45:22this repo over here. So, I'm going to go
45:24ahead back over to skills.
45:27And we are going to copy this here. I'll
45:31do SSH. That's what I prefer. And um I
45:35want to bring on down those skills. In
45:37fact, this folder is skills. So I might
45:40just want to
45:42back out for a sec. I'm just going to go
45:44my home directory. I'm just going to go
45:45get clone and then clone this
45:46repository. Okay. Because I don't want
45:49all of this stuff.
45:51Okay. And so I just want to move in.
45:53We'll go cd skills. I just want to move
45:55one thing. I want to move the
45:59um
46:03skill creator
46:05over into
46:08our claw directory into our skills
46:11directory. Okay. And then we'll go back
46:14over here. We'll make sure if it's
46:15there. So now we have that skill. Now it
46:17should be globally available. So we can
46:19do this wherever we want. And I mean
46:21really we shouldn't be really doing
46:22stuff in this folder. Um, so I'm just
46:24going to go back over to here and I have
46:26any kind of just go to any kind of
46:28project. Doesn't really matter.
46:30And um, go like another link tree. It
46:33doesn't matter. And we'll go ahead and
46:34type this type this in. And so now that
46:36skill should be available. So go skill
46:37creator. And so I want I want a skill
46:42that will uh, let that will back up my
46:47um, sessions for a for the current
46:51project. I am into a backup
46:56directory.
46:58The backups
47:02should be located at claude
47:08backups
47:10project.
47:12Okay. Um the other thing is that you
47:14know we might even want this to run on a
47:16schedule. Um, so that could be something
47:19that you might suggested to do, but you
47:22could also just trigger this manually.
47:23So I want to I want a skill that will
47:24back up my sessions for the current
47:26project. Uh, and the backup should be
47:28located at this. So docloud backups
47:31project. Um, you know, session files
47:34session files,
47:36you know,
47:38um,
47:40I think that's enough for it to know.
47:44Uh, it doesn't know what sessions we're
47:45talking about.
47:50Claude code sessions. Okay. And so you
47:54can have a skill in here in in here, but
47:57it probably will try to put it in the
47:59skills directory where this other one
48:00is. I notice it likes to default there,
48:02which is fine, but sometimes you want to
48:03have skills that are project specific.
48:05But other times you want them to go to a
48:07specific location. We'll say yes, allow.
48:10Yes, allow.
48:11And we will let it go ahead and create
48:13that skill.
48:15still has a few more things it wants us
48:16to accept. Yes, allow.
48:19And we'll give it a moment here. And
48:21here, notice it's putting in the cloud
48:22directory as I suspected. So, that's no
48:25surprise. Um,
48:28and it shouldn't take too long for this
48:30to figure it out.
48:36I think it's trying to test it now. I'll
48:37go ahead and say yes. Create that
48:39directory. And so the idea here is that
48:41it's not really hard to create a skill
48:42and it's also really not hard to uh uh
48:45create a backup. It's literally doing a
48:47backup right now. It's testing out the
48:48skill.
48:52Oh, it's actually running the iteration.
48:54And so I I think we talked about I'm
48:56going to stop this for a second. I don't
48:57want it to do that. Um but one thing
48:59that it can do is it can run evaluation.
49:02So I I walked us through like the agent
49:05skill implementation. It turns out that
49:07um uh when you create a skill, it has
49:12all that functionality built in. So you
49:13don't have to roll our own, which was
49:14something that I thought that we'd have
49:15to do. Uh but we don't. The only thing
49:18the reason I stopped is that it tends to
49:20consume a lot of credits. And unless you
49:24really need to test it, I try to avoid
49:27uh doing that. So now what I'm going to
49:28do is I'm going to go ahead back into
49:29Claude because I think the skill does
49:31exist now. And if I go here and just
49:33type in um is it backup session backup,
49:38it should try to now back up that
49:39session.
49:41So see here, it wants access to the
49:43projects directory
49:45and then it should move the files over.
49:49We'll say yes, allow it to do that.
49:53These might be things that we can
49:55specify in our skill to make it a little
49:56bit easier. And here you can see it's
49:58moving the files over. Okay. Okay. And
50:00so we've seen our backup and there's our
50:01backup. If we wanted to take this
50:03further, I would probably um maybe
50:07update the skill and say, "Hey, can you
50:09set this up on a cron job and can we
50:11edit the crown job and see whether it's
50:13running or not?" And that's what I would
50:14probably do. I'm not going to do that
50:16because I don't really want to back up
50:17my my stuff. If people have an interest
50:19with that, I could take it a bit
50:20further. But basically you go back to
50:21skill creator and just say like hey you
50:23know can you can you update my um backup
50:28uh session skill so that it can manage
50:33the cron job of my uh Linux machine here
50:37and run um and back up all projects uh
50:42whatever whatever right and you know it
50:44just specified that it has the ability
50:46to manage the cron job and you can ask
50:47it to you know list stuff or maybe have
50:50log We might also want like a log file
50:51to say back it up in certain places. So
50:53nothing super complicated, but you know,
50:55we got some experience with skills. Um,
50:57we might want to just take a look at
50:58that skill really quickly. So I'm going
51:00to go uh just quit out of here for a
51:02second and I'm just going to say code
51:05claude here. And we'll just take a look
51:07at this generated code in case I haven't
51:09covered skill creation. I just don't
51:11have a sense of it because I've been um
51:15uh doing a lot of doing a lot of
51:16development on my regular machine. I
51:18can't tell. We actually have this
51:19folder. This is actually just to run the
51:21eval. So I'm going to get rid of it.
51:22It's completely useless. And we'll open
51:24it up. And so here we have our
51:25skills.mmd.
51:27Okay. So it's telling us how the backup
51:29should work. And here is the eval.
51:33So like how should it evaluate whether
51:34it worked or not? I'm just going to go
51:35ahead and delete that. Um the only thing
51:37that I might look into is like okay, can
51:40we set the common things that it needs
51:42to set here? Um so I'm going to go over
51:44to our one second here. skill MD uh
51:48skill uh skills claude code. What I'm
51:51looking for is because I keep forgetting
51:53if this is an option, but if we can just
51:54set the allowed tools for the skill. I'm
51:57pretty sure we can set allowed skills
51:59for the tool.
52:02So say
52:04allow
52:07tools. There is.
52:11So I'll go back over to here.
52:15skill creator.
52:17Can you update
52:21my backup skill to have allowed tools
52:25for the the actions it will need? For
52:30example, it should be able to
52:33list and read the
52:38cloud projects
52:43backup directory
52:46and ex and etc. Okay. And so let's just
52:49see if it can do that because I don't
52:50want to have to keep asking or like
52:52accepting for very common stuff that we
52:54know that it needs. So hopefully it
52:56figures that out. But we'll give it a
52:57moment here to do that.
53:01See, it's like, yes, can I read sessions
53:03backup? Yes, of course you can. Um,
53:09and
53:13yes. [laughter]
53:15And so in that skill,
53:18yes. Okay, we're going to want to see
53:23that information. And I have the other
53:24one open over here. So, we'll be able to
53:25check in just a moment.
53:32Yes.
53:38Yeah. But we don't need a settings dot,
53:39Jason.
53:44Uhhuh. Just added to the skill.md.
53:48Maybe it doesn't know.
53:56It looks like there it's going to try
53:57it. Maybe I should have just told it to
53:59go to the file. There we go. Yep. Yep.
54:01Yep. Yep.
54:05It's going to edit it.
54:08A lot of this is just managing these
54:09darn allowed tools. They're really
54:10annoying.
54:12Allow. Allow. Yep.
54:17Really isn't that complicated. It's
54:19really taking forever here.
54:27Uhhuh.
54:31It's a thing you can do. Come on, buddy.
54:33Buddy, come on. Just write it. This is
54:35insane. I could have written this by
54:37now.
54:43What is it? Okay. What are you doing?
54:47Why aren't you just writing
54:51the tool used to the skills session
54:55backup
54:57uh skill.md
54:59and adding
55:02allow tools because like it like
55:05literally I'm just watching it just
55:07trigger all these darn commands.
55:12Oh, okay. There we go. Freaking took
55:15forever.
55:18There we are. And so here we have this.
55:21I mean it's not this is not good.
55:24Okay. But you know you need to read read
55:28and write certain files.
55:31This is too broad of permissions.
55:35Come on.
55:38That's not smart because then you're
55:39giving it like way too much um control
55:41to do whatever it wants.
55:48I should have told it to give it make it
55:49a plan first.
55:55Uhhuh.
55:57There we go. So, projects backups.
56:00There. That's what I want. And so, we
56:02could test it. I'm not really worried
56:03about testing it right now. We also
56:05might want the read and the right there.
56:06They don't have those as well, like the
56:08read and the right. Um, well, at least
56:10the read. But anyway, so that that's
56:13what I would call done. And I'll see you
56:15in the next one.

About this transcript

This page contains the full transcript of Claude Code Subagents and Multi Agent Workflows by ExamPro, generated from the public captions YouTube serves with the video. The transcript has 8,225 words across 1,267 segments, with the original timestamps preserved so you can click any line to jump to that moment in the embedded player.

What you can do with it

Use the transcript to take notes, quote the speaker, build a study guide, generate a summary with ChatGPT or Claude via the YouTube Summary tool, or export it as a timed subtitle file with YouTube to SRT. You can also re-open it in the transcriber to translate the transcript into 100+ languages.

Free YouTube transcript tool

YouTube2Text is a free YouTube transcript generator — no signup, no daily limit. Paste any YouTube link and get the full transcript instantly, with timestamps, click-to-jump, translation to 100+ languages, AI prompts for ChatGPT, Claude, and Gemini, and exports to TXT, SRT, VTT, or Markdown.