AI Summary
This tutorial demonstrates how to build a parallel AI agent workflow using Hermes, enabling multiple agents to research different companies simultaneously and compile a unified report. The key innovation is using a single agent profile for a role, rather than creating separate agents for each company, which allows the system to scale dynamically by simply adding tasks.
Chapters
One AI agent handles one task at a time, but by splitting a job into pieces and running one agent on each piece simultaneously, you achieve parallel processing. The same agent role (e.g., researcher) can be assigned to multiple different targets (companies) without creating new agent profiles.
A common mistake is creating four separate agent profiles for four companies. The correct approach is to create one agent profile for the role (e.g., 'researcher') and assign tasks for each company to that same profile. This scales: adding more competitors only requires adding more tasks, not new profiles.
Create a single profile named 'researcher' with a simple description: 'Researches a company's pricing and features.' Use a fast, cheap model like DeepSeek-V4-Flash for fetching and summarizing; the expensive model is reserved for the synthesizer that compiles the final report.
To enable the agent to create tasks from a Telegram message, the Kanban tool must be added. This involves backing up the config file, running a command to add Kanban to the tool set, then restarting the gateway. A conflict between two gateways can be fixed by asking Hermes to fix it via chat.
Create four empty tasks (one per company) in Kanban, then a fifth task for the synthesizer. The synthesizer task must have all four research tasks as its parents, so it only launches after all research is complete. The synthesizer prompt is crucial: it provides goals and output format (cover page, executive summary, charts, comparison table) but not step-by-step instructions.
The synthesizer waits for all parent tasks to finish before launching. Researchers are assigned to their respective tasks with company-specific prompts. The system automatically nudges tasks to start, and the parallel execution is faster than sequential human work.
The generated PDF includes a cover page, executive summary with a verdict ('ClickUp wins on price-to-feature ratio'), charts in brand colors, a full breakdown of pricing tiers, top features ranking, and a positioning matrix. The agent didn't just collect data—it made analytical judgments.
To automate daily runs, store the competitor list and builder instructions in the agent's memory via Telegram. Set up a cron job that pings the user every morning to confirm, then automatically creates the research tasks and builder task. Adding a new competitor is as simple as sending one chat message to update memory.
Adding 'Linear' to the competitor list via chat updates memory without triggering research. When the user says 'Go', the system dynamically creates five research tasks (one per company) and the synthesizer, all without manual reconfiguration. The workflow reshapes itself around a single word added to a chat.
By setting up a single agent profile for the researcher role and using Kanban to orchestrate tasks, you can build a parallel agent workflow that scales dynamically. The entire process runs automatically every morning, delivering a polished competitive analysis report to your phone with zero manual intervention.
Clickbait Check
90% Legit"The demo delivers exactly what the title promises: a complete parallel agent workflow built in around 15 minutes with no coding."
Mentioned in this Video
Tutorial Checklist
💡 Key Takeaways
Agent Role vs Topic
Clearly distinguishes the correct architectural pattern: one agent profile per role, not per topic, which is a foundational insight for scaling AI agents.
1:08Model Selection Strategy
Demonstrates cost optimization by using a cheap model for simple fetch-and-summarize tasks and reserving an expensive model for synthesis.
2:05Goal-Oriented Prompting
Shows how to write a prompt that gives the goal and output format, not step-by-step instructions, making the agent reliable and adaptable.
6:10Agent Taking a Verdict
Highlights that the agent didn't just collect numbers but made an analytical judgment, which is a key differentiator from simple data scraping.
9:36Dynamic Workflow Reshaping
Illustrates that adding a new competitor via chat automatically reshapes the entire workflow without manual reconfiguration, demonstrating true agent autonomy.
14:29Full Transcript
[00:00] One AI agent, one task at a time. That's
[00:02] the ceiling, and it's slow. So, we break
[00:04] it with Hermes. Run a whole team of
[00:06] agents in parallel. So, you take one
[00:08] job, split it into pieces, and put one
[00:11] agent on each, all working at the exact
[00:13] same time. Same role, different target.
[00:16] Five agents, five companies all at once.
[00:19] And you're never capped at five. Need a
[00:21] 10th next week? Text the bot one line,
[00:23] and it's in. No setup, no new agents.
[00:26] Then one last agent reads everything
[00:28] they found and turns it into a single
[00:30] report. And the whole thing is
[00:32] automated. It runs every morning on its
[00:34] own before you're even up. This is
[00:37] what's on my phone when I wake up, a
[00:39] full professional report. As you can
[00:41] see, five competitors built and
[00:43] delivered overnight. I didn't touch it.
[00:46] No code.
[00:47] Let me show you how to build this from
[00:49] scratch.
[00:54] So, we've got Kanban open.
[00:56] We're building four agents that research
[00:59] for companies at the same time. Then one
[01:01] that turns it all into a report. So,
[01:03] first, let's click on profiles,
[01:06] where you set up your agents.
[01:08] So, here's how I did it the first time.
[01:10] Four agents, Asana, ClickUp, Monday, and
[01:14] Notion. Four companies, four agents. And
[01:16] that's wrong because an agent is a role,
[01:19] not a topic. A researcher is a
[01:22] researcher whether it reads Asana or
[01:24] Notion. Same job, you're just pointing
[01:27] it at a different company.
[01:29] It's like hiring four people because
[01:31] you've got four reports to write. You
[01:33] don't. You hire one researcher and hand
[01:36] them all four. And one researcher
[01:39] scales. 10 competitors next week? Same
[01:42] agent, more tasks. If you have four
[01:44] profiles, you'd build a new one every
[01:47] single time, and it's a dead end. So,
[01:49] this go. Let's delete them.
[01:53] And of course, we keep default. And now,
[01:55] the right way. One profile. Let's click
[01:58] on create.
[01:59] Let's name it
[02:01] researcher.
[02:02] For the description, let's keep it
[02:04] simple.
[02:05] Researches a company's pricing and
[02:08] features. And that's the role stated
[02:11] once. Doesn't matter which company.
[02:13] That's the whole point. The model is
[02:16] DeepSeek-V4-Flash,
[02:17] not the pro one, because these agents
[02:20] just fetch and summarize. Flash is fast,
[02:23] cheap, plenty for reading a pricing
[02:26] page. The smart model comes in later,
[02:29] where it earns it.
[02:30] So, DeepSeek-V4-Flash and create. So, we
[02:34] have one researcher, and that's the
[02:36] agent doing all four jobs.
[02:39] So, one more thing before we built.
[02:41] Right now, the default agent runs the
[02:43] bot, but it can't create tasks on its
[02:46] own from a Telegram message, and we need
[02:49] that at the end. So, we add the tool
[02:51] now. Let's open our terminal, and let's
[02:53] copy this command. So, this backs up the
[02:56] config file, a copy, in case we break
[02:59] something editing it. Anything goes
[03:01] wrong, we restore from this. So, let's
[03:03] run it. No output means it worked.
[03:07] Okay, great. Our next command. Let's
[03:10] paste it. So, this one adds Kanban to
[03:13] the tool set list. The line that gives
[03:15] the agent the power to create tasks.
[03:17] Let's run it.
[03:19] Great. Now, let's check.
[03:22] And here we go. We added Kanban. So, we
[03:25] have Hermes-CLI
[03:27] and Kanban, both there. Now, let's
[03:30] restart the gateway.
[03:32] This will reboot the agent, so it loads
[03:34] that new Kanban tool. Let's run it.
[03:38] Okay.
[03:39] Gateway is shutting down.
[03:41] So, now we can see that start has
[03:44] failed. Let's fix that. First, let's
[03:48] type anything to Telegram, and it's
[03:51] working.
[03:52] But, we still have start failed. Let's
[03:54] click on it.
[03:56] Okay. So, the easiest way to fix this is
[03:59] to copy this.
[04:02] I know it sounds very lazy, but let's go
[04:04] to chat.
[04:06] Let's paste this here and tell Hermes to
[04:09] fix it.
[04:11] And we'll wait.
[04:15] Ah, here we go.
[04:16] It's running right now. I think it's
[04:19] fixed. It's usually because it's a
[04:21] Telegram conflict, two gateways fighting
[04:24] over the same bot.
[04:25] So, by telling Hermes to fix it, it will
[04:28] close all the gateways and keep defaults
[04:30] running.
[04:34] Ah, here we go. Everything is fixed. You
[04:37] can see the root cause. The researcher
[04:39] profile's gateway was holding the
[04:41] Telegram bot token, blocking the default
[04:44] profile's gateway from starting. For
[04:46] this kind of problems, it's better to
[04:48] tell Hermes to fix it itself.
[04:50] So, [snorts] setup's done. Now, we pawn
[04:52] this one agent at four companies and let
[04:55] it run. So, let's head over to Kanban.
[04:58] Our dashboard. So, first thing to do is
[05:01] to create the tasks, one per company.
[05:04] So, first one is ClickUp.
[05:08] Let's create, and then we have Monday.
[05:13] Then Asana.
[05:16] And lastly, Notion.
[05:20] Okay. Four tasks, empty for now, no
[05:23] assignee, no description, nothing. They
[05:25] just sit there.
[05:27] So, now the one that matters, which is
[05:29] task five, the agent that pulls
[05:31] everything together.
[05:33] So, let's add it.
[05:35] The name is the synthesizer. Let's
[05:37] assign it to default.
[05:41] And for parents, all of these tasks are
[05:44] its parents. So, let's select one for
[05:47] now. Let's create and let's open it up
[05:51] because this one needs real prompt.
[05:53] First thing's first, let's add more
[05:56] parents, which is Monday,
[05:59] Asana, and Notion.
[06:01] And for the description, let me copy my
[06:04] prompt and let's explain it.
[06:07] So, this is the prompt that matters and
[06:10] I want you to notice how it's written.
[06:12] I'm not telling it which buttons to
[06:14] press. I'm giving it the goal and the
[06:16] output. Read what the four researchers
[06:18] found, build a report, here's what the
[06:21] report should look like. The design,
[06:23] that's my brand colors, dark background,
[06:27] gold fonts. I'm adding it a style guide.
[06:30] The content, a cover, a takeaway, line
[06:33] up top so it makes a point instead of
[06:36] just dumping data, charts, a comparison
[06:39] table, top five features. And the last
[06:41] line, the media one, that's what drops
[06:44] the finished file straight onto my
[06:46] phone. What I'm not doing is telling it
[06:48] how to install anything or which library
[06:51] to use. I give it the what, it figures
[06:53] out the how. That's what makes an agent
[06:56] reliable instead of brittle. Okay? So,
[06:59] we are done here. Let's save and let's
[07:02] go back to our dashboard. And watch,
[07:05] even with default assigned, the
[07:07] Synthesizer does not launch. It just
[07:09] sits because it's waiting on all four
[07:11] parents to finish first. That's the
[07:14] funneling. One agent holding until
[07:16] everything it needs is done. Now, let's
[07:19] arm the researchers. Each one gets the
[07:22] same agent, researcher, and the same
[07:24] instructions pointed at a different
[07:27] company. So, let's start with ClickUp.
[07:30] So, assignee is researcher
[07:33] and let's edit the description. Save.
[07:38] Second one is Monday, so researcher.
[07:41] Let's edit it.
[07:45] So, for ClickUp here, after
[07:47] researcher, let's hit enter. Now, it's
[07:50] done.
[07:51] Asana
[07:52] same thing
[07:55] Let's edit the description.
[07:58] And Notion So, researcher
[08:02] And let's paste our command, our prompt,
[08:06] and click save. Now, they start on their
[08:08] own in about a minute.
[08:11] The bot checks every 60 seconds, but I
[08:13] don't want to stare at it. So, let's
[08:16] click on nudge. And here we go. They're
[08:18] in progress. So, the first one is
[08:21] already done. Let's wait for the other
[08:23] three.
[08:24] So, while these run, here's the thing
[08:26] people miss about agents. They're not
[08:28] faster because they type faster, they're
[08:30] faster because they work at the same
[08:32] time. One of me does these four
[08:34] back-to-back. Four agents do them at
[08:37] once.
[08:38] And they're finishing.
[08:40] So, all four done, and now task five
[08:42] should wake up. Let's scroll here.
[08:47] So, it's in ready. Let's click on nudge.
[08:50] And now it's in progress, and let's wait
[08:53] for it.
[08:54] It will be reading all four reports and
[08:56] building the PDF.
[08:58] >> [snorts]
[08:58] >> First run, it installs the rendering
[09:00] library, so it takes a minute.
[09:03] And then the agent will solve its own
[09:05] setup.
[09:06] So, here, nothing for me to do. We'll
[09:08] wait for the Telegram message.
[09:11] And there it is, straight to Telegram. I
[09:14] didn't write this, didn't design a
[09:16] single page for agent research. One
[09:19] agent built it, and it landed on my
[09:21] phone as a finished file.
[09:23] So, let's open it.
[09:25] And look at this. A cover page, title,
[09:28] the four platforms, the quarter, the
[09:31] date. This isn't a text dump. It's a
[09:33] document. And right here, page two, the
[09:36] executive summary.
[09:38] Read that first line, ClickUp wins on
[09:40] price-to-feature ratio.
[09:42] That's not data, that's a verdict. The
[09:45] agent didn't just collect numbers, it
[09:47] took a position.
[09:48] I come back to that because that's the
[09:50] most important thing in the whole
[09:52] report. And the chart, two of them,
[09:54] entry tier and most popular tier, side
[09:57] by side in my colors, the gold, the dark
[10:00] background, it built those from the
[10:02] numbers the researchers built. Then we
[10:05] have the full breakdown, every tier,
[10:08] every price for all four, ClickUp,
[10:10] Asana, monday, Notion, what each plan
[10:14] costs and what you actually get. And you
[10:17] have here top five features across
[10:19] platform ranked, AI agents, automations,
[10:22] the view types, integrations, enterprise
[10:25] security one to five. On the last page,
[10:28] this is the one, a positioning matrix.
[10:31] Each platform, its price range, its
[10:34] wrist strength, what it's for, ClickUp
[10:36] for cost-conscious teams, Asana for
[10:39] regulated enterprises. This is strategy,
[10:42] not a spreadsheet. So, for agents
[10:44] research this, something you'd actually
[10:47] send a client. But there is one more
[10:49] thing I still had to do. Sit here and
[10:52] start it. So, let's fix that. Let's make
[10:54] it an automated workflow where you can
[10:57] add competitors and set that to work
[11:00] every day.
[11:02] So, to do that, let's go first to
[11:04] Telegram. So, the first thing to do is
[11:07] to give the bot a memory, my competitors
[11:09] once right here in the chat. So, let's
[11:12] paste this command. Okay, memory my
[11:15] competitor list, ClickUp, monday, Asana,
[11:17] Notion. Remember this, I'll ask you to
[11:19] run research on them later. Let's run
[11:21] it.
[11:24] And now it knows my list and I can
[11:26] change it anytime just by talking to it.
[11:29] You'll see. So, one more thing in its
[11:31] memory, the builder's instructions. So,
[11:34] when it runs, it knows exactly how to
[11:36] assemble the report. This is the brain
[11:38] behind that PDF. Let's paste this, and
[11:42] let's run it.
[11:46] So, now we have the list and the builder
[11:48] both in memory. Now, the schedule. I
[11:50] want it to ping me every morning and ask
[11:53] if I want to run. And here's the catch,
[11:55] a scheduled job can't ask me and then
[11:58] sit there waiting for an answer. It runs
[12:00] alone, no chat. So, it pings me and I
[12:03] reply separately. Ping, then reply. So,
[12:07] let's paste our cron command, and let's
[12:10] run it. And that's the whole engine in
[12:13] one instruction. The ping only asks, but
[12:16] go builds everything. A research task
[12:18] per company and the builder on top
[12:21] waiting on all of them. And I made it
[12:23] confirm the count back to me, so I
[12:25] always know the builder's there. So, I'm
[12:28] not waiting until 9:00 a.m. tomorrow to
[12:29] show you. I'll tell it to run the job
[12:32] right now. So, let's say, "Run the daily
[12:35] search ping job now." And let's run it.
[12:39] So, let's wait for it. It's running, and
[12:43] there's the ping, "Morning. Here's your
[12:45] list ClickUp Monday Asana Notion.
[12:48] Edit or say go." And it pulled that from
[12:51] memory. I didn't paste it. So, let's
[12:53] say, "Go." And then enter.
[12:56] And it confirms. So, it's creating the
[12:59] four tasks, as you can see here. And
[13:02] then the synthetizer. So, let's go to
[13:05] the board. Here you go, four researchers
[13:07] and one synthetizer. Great.
[13:11] So, now let's wait for our PDF, and
[13:13] after that we run a new competitor.
[13:17] Great. So, everything is working. Now,
[13:20] the part that makes this useful, the
[13:23] dynamic workflow, because our list isn't
[13:26] locked. When a new competitor shows up,
[13:28] we just tell it to add it. For example,
[13:32] let's add Linear to the list.
[13:34] So, add Linear to my competitor list
[13:37] just update the memory list. Do not
[13:39] create or run anything yet. Let's run
[13:41] it.
[13:43] And it just updates the memory. Nothing
[13:45] runs. Adding a name and running the
[13:47] research are two different things. It
[13:50] knows the difference. Lists, we have
[13:52] five competitors now.
[13:54] And the board hasn't moved.
[13:56] Now, let's say go.
[13:59] And hit enter.
[14:01] So, five research tasks now and the
[14:03] builder on top of all five. So, I have
[14:06] six.
[14:09] So, let's see the board. So, we have
[14:12] five research tasks and the synthesizer,
[14:15] the sixth one.
[14:17] So, five researchers, one for each
[14:19] company including Linear, and the
[14:21] builder waiting on all five. Great.
[14:24] So, I added one word to a chat message
[14:26] and the whole machine reshaped around
[14:28] it.
[14:29] That's an agent, not a script. I didn't
[14:32] rewire anything. I told it.
[14:34] So, here is where we landed. Every
[14:36] morning at 9:00, it pings me with my
[14:38] list. I reply, edit it, or just say go.
[14:42] And it builds the whole thing. Finds
[14:44] out, researches everything, sends me the
[14:46] finished report.
[14:49] I set it up once. Now, it runs itself.
[14:55] And here's the final report. Same
[14:57] engine, but now with Linear in it. Let's
[15:00] click on it.
[15:02] And right on the cover, we have ClickUp,
[15:04] Asana, Monday, Notion, and Linear.
[15:07] If you scroll down,
[15:09] we can see Linear here as part of the
[15:11] chart. So, it didn't fill a slot. It
[15:13] researched Linear and took a position on
[15:16] it. That's the whole thing. Five
[15:18] competitors, one word added to a chat,
[15:21] four reports on my phone, and tomorrow
[15:23] at 9:00 it asks me again.
[15:26] And that's parallel agents in Hermes.
[15:28] We're all many targets running on their
[15:31] own every morning. If you want to run
[15:33] Hermes 24/7, there is a VPS link in the
[15:36] description. Use the code nocodehack for
[15:39] a discount. Like the video, subscribe
[15:41] for more Hermes builds, and I'll see you
[15:43] in the next one.