Speaker A
Most people prompt Claude the same thing over and over again because they don't know how to use agents yet. With agents in Claude Cowork, we can now bulk process dozens of tasks in parallel and run full multi-step workflows autonomously.
Learn how to use agents in Claude Cowork to bulk process leads in parallel, build skills and commands, and automate multi-step sales workflows from scratch.
Key Takeaways
- Parallel sub-agents in Claude Cowork dramatically speed up bulk processing of tasks like lead qualification and outreach
- Structuring work into skills, commands, and agents turns repeated prompting into reusable automated pipelines
- Explicitly instructing Claude to use parallel sub-agents is necessary to avoid slow sequential execution
What the video covers
- Most users repeat the same prompts in Claude because they haven't learned to use agents yet
- Claude Cowork agents can process dozens of tasks in parallel and run multi-step workflows autonomously
- A demo qualifies 150 marketing agency leads using 15 parallel sub-agents, identifying 82 as qualified in about 2 minutes
- Qualified leads are enriched with LinkedIn and company data via 18 sub-agents, including a LinkedIn scraper
- Personalized icebreaker outreach messages are then generated by 17 agents and added to the CSV sheet
- Claude plugins combine skills, commands, agents, and connectors to automate repeated and complex tasks
- Skills save specific ways of working; commands string skills together; connectors link Claude to third-party software via APIs or MCP
- Sub-agents keep the main agent's context lean by returning only summaries, improving performance and enabling parallel work up to roughly 100x faster for suitable tasks
- Prompts must explicitly request parallel sub-agents, since otherwise Claude may run them sequentially and slow the process
- Teams work better for tasks needing shared coordination, while sub-agents suit tasks that can be done in isolation, such as most sales, marketing, and support work
Chapters
- 00:00Introduction to Agents in Claude Cowork
- 01:02Demo: Qualifying and Enriching Leads with Sub-Agents
- 03:24Combining Skills into an Outbound Pipeline Command
- 04:50How Plugins, Skills, Commands, and Connectors Fit Together
- 06:07Why Sub-Agents Improve Performance and Context Use
- 07:28Prompting Claude to Use Parallel Sub-Agents
- 08:28Building a Sales Plugin from Scratch with Apify Connectors
- 12:54Wrap-Up: Structuring Agents, Skills, and Commands
Full Transcript — Download SRT & Markdown
Speaker A
autonomously. So, in this video, I'll show you how agents work in Claude Co-work, show you demos and examples of how to use them, and show you how to set them up from scratch to automate your workflows. Now, if you're new to Claude
Speaker A
So, in this video, I'll show you how agents work in Claude Cowork, show you demos and examples of how to use them, and show you how to set them up from scratch to automate your workflows.
Speaker A
know yet is how to use agents. Using agents is a little bit more advanced, but even if you're new to Co-work, you should be able to follow this. If not, I recently launched a full beginner tutorial, which I'll make sure to link
Speaker A
Now, if you're new to Claude Cowork, Cowork is a game-changer because we can now give Claude access to files on our computer, connect it to our entire software stack, and also create skills to automate some of our workflows. Now, what many people don't know yet is how to use agents. Using agents is a little bit more advanced, but even if you're new to Cowork, you should be able to follow this. If not, I recently launched a full beginner tutorial, which I'll make sure to link in the description below, too.
Speaker A
For example, here I gave Claude a lead list of more than 150 marketing agencies, which we got from Apollo, a lead database, and then I let it qualify the list with sub agents first according to our ICP criteria, which is a
Speaker A
Now, before explaining exactly how these agents work, let me show you a quick example of what they can be used for.
Speaker A
launched 15 sub agents in parallel and then identified that only 82 were qualified. And because these sub agents work in parallel, this only took 2 minutes. After that, it went to the next step in this workflow, which is
Speaker A
For example, here I gave Claude a lead list of more than 150 marketing agencies, which we got from Apollo, a lead database, and then I let it qualify the list with sub-agents first according to our ICP criteria, which is a marketing agency providing SEO services and located in the US. And then you can see it launched 15 sub-agents to qualify all 150 leads in parallel. Each agent will use web search to verify the company actually offers SEO services and is US-based.
Speaker A
personalized outreach messages based on the lead intelligence data. The personalized outreach messages are also added to the sheet. So, as you can see here, in the same chat in Claude, it qualified 82 out of 150 leads, it enriched 82 leads, and then wrote 37
Speaker A
So, here it launched 15 sub-agents in parallel and then identified that only 82 were qualified. And because these sub-agents work in parallel, this only took 2 minutes.
Speaker A
of these skills can be strung together and combined into an agentic workflow through commands, and we can even save the specific agent instructions in the agent section. So, you can see in our sales plugin here, we have specific
Speaker A
After that, it went to the next step in this workflow, which is enriching the leads with more information. So, it launched another 18 sub-agents: 17 lead researchers and one LinkedIn scraper. It then added that qualification status and lead intelligence to the CSV sheet, and then it spun up another 17 agents to write personalized outreach messages based on the lead intelligence data. The personalized outreach messages are also added to the sheet.
Speaker A
process it has to go through, and instructions on how and when to use a sub agent here. The instructions for what these sub agents have to do are saved here in another tab, which is called agents.
Speaker A
So, as you can see here, in the same chat in Claude, it qualified 82 out of 150 leads, it enriched 82 leads, and then wrote 37 icebreaker messages. And this is really the power of agents, because this was very difficult to do before we could use sub-agents.
Speaker A
with all the agents with just one slash command. So, you can see that command we set up for outbound pipeline is just a sequence of steps to take and skills to use in order to execute a more complex
Speaker A
Now that we have plugins, each one of these bulk tasks can be saved as skills. Multiple of these skills can be strung together and combined into an agentic workflow through commands, and we can even save the specific agent instructions in the agents section. So, you can see in our sales plugin here, we have specific skills for each of the tasks in the workflow: lead qualification, lead intelligence, and email personalization, with specific instructions on how we exactly do these processes. For example, in the lead qualification skill, we have our ICP criteria, the qualification process it has to go through, and instructions on how and when to use a sub-agent. The instructions for what these sub-agents have to do are saved in another tab, which is called agents. Here, for example, we have the lead qualifier sub-agent with the specific instructions on what it needs to do.
Speaker A
you, but it's not, and I'm going to show you how to set this up yourself from scratch. If you want access to all of the skills and plugins that I just showed, plus all of the other skills and
Speaker A
We've combined these three skills into one command to trigger all of these skills in sequence, together with all the agents, with just one slash command. That command, outbound pipeline, is just a sequence of steps and skills to use in order to execute a more complex workflow. This allows us to run this multi-step workflow with sub-agents with a single command, which I can trigger simply by going to plugins, sales, and outbound pipeline.
Speaker A
if that's interesting to you, definitely check it out in the link in the description. Now, before showing you exactly how to do this yourself, let me quickly explain how these sub agents work and why it opens up uh lots of new
Speaker A
Now, this might look a little complicated to you, but it's not, and I'm going to show you how to set this up yourself from scratch. If you want access to all of the skills and plugins that I just showed, plus all of the other skills and plugins we're building out, you can check out my AI accelerator, where we list all our skills. We also do weekly AI workshops, where we dive much deeper into these tools, and we have one-on-one unlimited live tech help available. So if that's interesting to you, check out the link in the description.
Speaker A
recently did a full uh video on it, which if you haven't seen it, I highly recommend checking it out. I'll make sure to link it in the description below, too. But, essentially, through skills, we can save specific ways of
Speaker A
Now, before showing you exactly how to do this yourself, let me quickly explain how these sub-agents work and why they open up lots of new possibilities for Claude Cowork, and also tell you what the limitations still are.
Speaker A
And before agents were available in Claude, bulk processing was difficult for many tasks because we essentially only had one agent to process each task separately. This overloaded the context window, made it extremely slow, and often just not viable to do for these
Speaker A
With Claude plugins, which were recently launched, we now have skills, commands, agents, and connectors available in Claude. If you don't know what skills are yet, I recently made a full video on them, which I highly recommend checking out. I'll link it in the description below too. Essentially, through skills, we can save specific ways of working in Claude to automate repeating tasks. With commands, we can string together multiple skills in sequence to automate more complex tasks, and connectors allow us to connect Claude to any third-party software through APIs or MCP.
Speaker A
intelligence sub agents uses a LinkedIn scraper to scrape uh LinkedIn data. So, if we tell Claude CoWork to research 100 leads, for example, he could spin up uh 10 sub agents, each researching 10 uh leads um through, let's say, the
Speaker A
We also now have agents. Before agents were available in Claude, bulk processing was difficult for many tasks because we essentially only had one agent to process each task separately. This overloaded the context window, made things extremely slow, and often made these bulk processes not viable.
Speaker A
one of the most precious resources for these agents, and therefore, it will drastically improve the performance, first of all. And because sub agents can run in parallel, a task like this one in the example can be performed theoretically 100 times faster. And
Speaker A
But Claude can now spin up sub-agents, which can process tasks in parallel, meaning at the same time. Each of these sub-agents can also use tools and skills. For example, our lead intelligence sub-agent uses a LinkedIn scraper to pull LinkedIn data. So if we tell Claude Cowork to research 100 leads, it could spin up 10 sub-agents, each researching 10 leads through the LinkedIn scraper and other tools, and then give a summary of the lead research back to the main agent.
Speaker A
more like an automation tool comparable to N8N or make.com. Now, we're not completely there yet, which I'll get to later in this video, but because this is now possible, hundreds of new use cases pop up for automation right inside
Speaker A
Why is this a big deal? Because the main agent won't be bloated with research context, just a summary. Context is one of the most precious resources for these agents, so this drastically improves performance. And because sub-agents can run in parallel, a task like this one can theoretically be performed around 100 times faster.
Speaker A
think of doing SEO audits. Of course, we can also combine these new use cases together to do autonomous multi-step automations in bulk, like I showed you in my example. Now, before showing you how to set it up and how to do this
Speaker A
Because of this, we can start using Claude Cowork for bulk task execution and deep research tasks, and also for autonomous multi-step workflows. These three things together make it much more powerful, and it becomes much more like an automation tool comparable to n8n or Make.com. We're not completely there yet, which I'll get to later in this video, but because this is now possible, hundreds of new use cases pop up for automation right inside Claude Cowork. You could create bulk ad variations, write bulk product or SEO descriptions, do bulk email personalization, run data analysis on thousands of transcripts, and pull all the research data back into a presentation. You could also run SEO audits. And of course, we can combine these use cases to build autonomous multi-step automations in bulk, like the example I showed you.
Speaker A
even run multi-agent teams that can communicate between each other. While Claude Co-work, they're all isolated sub-agents. Now, you can imagine that when they can communicate with each other, it really becomes an even more powerful setup because they can work
Speaker A
Before showing you how to set it up yourself, there's one more thing I want to clarify. Claude Code recently came out with a new multi-agent feature, but it's different from the sub-agents available in Claude Cowork. In Claude Code, we can now run multi-agent teams that communicate with each other, while in Claude Cowork, they're all isolated sub-agents. When they can communicate with each other, it really becomes a
Speaker A
large team of engineers to do for months. Now, this is really, really powerful, but I see it mostly as an engineering feature. Multi-agent teams are good for tasks where agents have to work together on one shared goal or uh
Speaker A
work together on a to-do list, while sub-agents are preferred when each one can work in isolation. And for most day-to-day office-related work and tasks in marketing, sales, customer support, etc., the sub-agent setup is usually enough and can actually be preferred
Speaker A
setup because it uses less tokens and it's really all we need for most of these isolated tasks like lead research, analyzing data, etc. Now, even though sub agents use less tokens than agent teams, they still use a lot of tokens.
Speaker A
So, you that's one thing you want to keep in mind and is still a limit limitation. It can use quite a lot of tokens when you spin up these sub agents. Of course, this going to get cheaper and cheaper, but if you want to
Speaker A
start doing this, you probably want to be on the Claude Max plan to not be hitting your limits all the time. And second, what I've noticed is for really large bulk tasks, this is probably not a setup you can use yet. I've seen this
Speaker A
work best for batches of 100 to 200. More than that, it's going to get complicated, take a long time, cost a lot of credits.
Speaker A
That's why for large bulk workflows, especially ones that are non-human in the loop, an automation platform like n8n or make.com is still a better option to automate those types of processes.
Speaker A
So, knowing that, let me show you how to use these agents in Claude Co-work. So, first, I'll show you how to make sure Claude uses these parallel sub agents because there is a specific way you need to prompt it. Then, I'll show you how to
Speaker A
save these tasks as skills and agent files. And then, I'll show you how you can build these commands and add them all into a plugin that you can easily share with your team or with other people. But, let's say we're going to
Speaker A
replicate that same use case as before, but in this case, I don't have the skills yet. Then, give it the CSV sheet of the leads, and then I would prompt it with something like this. I want to qualify these leads against my ICP
Speaker A
criteria. My ICP criteria are are that we only work with marketing agencies who provide SEO services and are based in the US.
Speaker A
Please spin up 15 parallel sub agents that each research 10 leads in this list and make sure they're qualified leads.
Speaker A
Based on the summaries you get back, you deliver an updated CSV sheet and add a column with the qualification status.
Speaker A
Now, specifically mentioning that it has to spin up these parallel sub agents is important because sometimes, if you don't say it, it will actually run sequential sub agents, meaning that they'll be running one after the other, which means it's going to take a long
Speaker A
time, and sometimes it won't even run sub agents if you don't specifically tell it to. So, you can see it already made a plan here, right, launching 15 parallel sub agents, collect results, and generate updated CSV.
Speaker A
So, you can see it's now spinning up these batches of sub agents to process and qualify all of the leads.
Speaker A
And you can see each of these sub agents does multiple queries on the internet on each of the leads to find more information about them.
Speaker A
So, it's now done the qualification on 150 leads, gave me a CSV sheet, which I can throw in through sheets, and now we have a new column with qualification status and a qualification reason column.
Speaker A
Now, if this is a repeated workflow that I would do, I could already save this as a skill by just telling Claude, "Please save this as a skill." So, next time I don't have to give all of these
Speaker A
instructions, it will already know exactly what to do with a lead list it will get. But, I'll show you one more example, and then we're going to build the entire plugin right away. So, we're going to build multiple skills,
Speaker A
commands, and agents into a plugin right away. So, let's say now we want to enrich the leads a bit, so we can use that enriched data later for personalized outreach. Now, for this task specifically, and generally if you're going to enrich leads, I'd highly
Speaker A
recommend installing the Apify MCP. Now, if you don't know Apify yet, Apify basically has thousands of scrapers, including scrapers for hard-to-scrape websites like LinkedIn, Instagram, Facebook, and through the Apify MCP, we basically get instant access to all of
Speaker A
these social media and websites that Claude cannot access. Now, you can just make an account for free here on Apify.
Speaker A
You get some credits, but if you're going to use it a lot, I recommend upgrading, and it will cost you $29 a month. Now, once you're in your account, you need to get your API key, which you can get here by going to settings and
Speaker A
API integrations, and here you can copy your API key. Then, in Claude, you can go to the connectors in the settings, and click connectors, and you go to browse connectors, and you look for Apify. You'll see that appear. Now, mine
Speaker A
is already connected. Um but once if it's not connected yet, uh it will ask you for the API key and enable tools, which basically specify which scrapers from Appify we're going to use. Now, the easiest way to do
Speaker A
this is by going to mcp.appify .com. And here you can basically specify all the scrapers you want to give access to Cloud.
Speaker A
So, first I'd recommend selecting all here. Then you can click here on add or remove actors. And here you can specify all the ones you want. Let's say, for example, LinkedIn mass LinkedIn profile scraper and let's say a LinkedIn post scraper.
Speaker A
This one, for example. So, we actually know the last posts that our leads have posted, so we can use that as a personalization. So, once I've selected, you can select way more. You click on save changes. Then you go here below
Speaker A
and you take this part after the equal sign. You copy that. You go back to Cloud and that's what you add here in the enable tools and you click on save.
Speaker A
Sometimes you have to restart your Cloud Co-work for Appify to actually be accessible, but then you should have Appify installed. And now I can say something like, "Please spin up 10 parallel sub agents that enrich seven qualified leads each. They should find
Speaker A
the leads' last LinkedIn post, the company size and the company description using the Appify scraper, specifically the LinkedIn scrapers to get these details." And then you should add these data points to the CSV sheet so we can later use that data for personalized
Speaker A
outreach. Again, I specified the use of parallel sub agents and also how many leads each sub agent should process. In general, what I've seen is the best is to have between five and 15 leads or or tasks per sub agent. Otherwise, it could take
Speaker A
a lot longer if you do more. And if you spun it spin up too many sub agents, it also runs into complications sometimes in my experience. So, you can see it spun up again 10 sub agents. You can see
Speaker A
each of the sub agents is using a LinkedIn profile scraper. So, it's now enriched the CSV with the last LinkedIn post if they had one, company description, and company size.
Speaker A
So, you can see we get the last LinkedIn post here. The enriched company size and the company overview.
Speaker A
Now, of course, now we can do another skill to write personalized outreach messages based on this enriched data, but I think you get the point. So, let me show you how you can now convert this into skills agent.docs, so you don't
Speaker A
have to repeat what these agents have to do, and commands to string multiple of these tasks together into one workflow.
Speaker A
The amazing thing in Cloud is that we can build these skills right from prompt. Now, because we're going to create multiple skills, we're going to create commands and agents, I would advise you to add them all into one
Speaker A
plugin. So, it's all stored inside of the same plugin, which then can also easily be shared. So, I could say something like, uh, for example, "Please create a new plugin that's called Ben's outbound, uh, sales process." Inside of that plugin, I want to create
Speaker A
one skill for the exact way we did the lead qualification and call it Ben's lead qualification skill.
Speaker A
Make sure to also create an agent.md, and the skill should refer to the agent.md instructions on what the agents need to do.
Speaker A
Also, create a skill for the exact way we did the enrichment process and call it Ben's enrichment skill.
Speaker A
In that one, also make sure to create a separate agent.md for the enrichment process. In both of these skills, it should clearly be instructed to spin up parallel batches of sub-agents, just like, uh, we did in the task here above.
Speaker A
Again, this is important to specify also when you're creating the skills or the plugin um because then it it's clearly stated in the skill to use those parallel, uh, sub-agents. Of course, we can also combine this into a command to trigger
Speaker A
both of these skills or tasks at once with one specific command. So, for example, I could say something like, "Also, please create a command.md where I can trigger both of these skills, uh, in sequence when I trigger a slash
Speaker A
command uh, with called outbound enrichment. This command, uh, should ask the user for a CSV sheet with leads, and it should then automatically use the lead qualification skill and the lead enrichment, uh, skill in sequence to deliver a fully enriched CSV to the user
Speaker A
at the end. All of these new documents and files should be stored in the same plugin.
Speaker A
So, you can see now starts building the plugin. It makes to-do list, click create plugin directory, create lead qualification skill, create lead enrichment skill, create outbound enrichment command.
Speaker A
So, now created the entire plugin with two skills, lead enrichment and lead qualification, one command, outbound enrichment, as you'll see, it already includes all the instructions on how to run this workflow, right? So, first use the lead qualification, then the lead enrichment.
Speaker A
We have the agents, the enrichment agent and the qualification agent. Now, sometimes it gives you a button which is called install, so you can install this as a plugin right away.
Speaker A
Sometimes it don't doesn't work, in my experience, so you can just ask it to give you the plugin as a zip file. This then you can use to share it to other people or to import it yourself if you
Speaker A
don't have the install button. And then you can just go to your plugin section, go to plus and click on upload plugin, and then you can drag that in, and upload.
Speaker A
And now you can see I have my new plugin with commands, skills, and agents.
Speaker A
You can also upload these plugins onto a GitHub account, so you can get a link, a versionable link that you can share with other people. And now that I have these commands and skills, I can of course trigger them anytime I want to do this
Speaker A
task again right inside Claude. So, I can just go and slash command, go to plugins, and trigger the outbound enrichment command. Now, that's it for this video.
Speaker A
Again, if you want access to all the plugins and skills we're building, uh, to maybe use or customize them for yourself, you can check out my AI accelerator. We also do weekly AI workshops where we dive a lot deeper
Speaker A
into these tools. We have one-on-one tech help unlimited to help you out in any tech problems you have and blueprints on how to start your own business. So, if that's interesting to you, you can check the link in the
Speaker A
description below. If you got any value out of this video, I really appreciate a like and a subscribe. It really does help me. If you want to learn more about Cloud Cowork, you can check out the video here above.
Topics:Claude CoworkAI agentssub-agentsparallel processingClaude pluginsagentic workflowlead qualificationsales automationskillscommands
