Hermes Agent Kanban + Cron Job is POWERFUL (Setup Guide)

All right, folks. I've been using the Hermes Kanban feature to improve my daily AI news cron job. So, if you're running cron jobs with Hermes agent and

you want better reliability, parallel execution, and proper error handling, that's the most important part, I'd say, then this video is for

you. I'm going to show you exactly how I

you. I'm going to show you exactly how I migrated from a single simple cron job. So, this is the output. Ignore the Ignore the fact that

output. Ignore the Ignore the fact that there's no, you know, wow visuals. I

just want it in markdown file. I just

need the, you know, the quality of the research itself. But, this is the before

research itself. But, this is the before of the AI news pipeline. This was

without the Kanban.

And this is after with the Kanban. You

can see noticeably it's a much more detailed.

It has more sources as well. So, I'm

going to show you how I use the Kanban feature to improve this pipeline. And

also show you what broke along the way and how I fixed it. But, as an important caveat here, it's not fully fixed.

There's still improvements needed to be done. But, I think that this is a very

done. But, I think that this is a very important feature that you really should be using to improve the pipelines of your projects. It doesn't even have to be

projects. It doesn't even have to be cron jobs. But, cron job is just one

cron jobs. But, cron job is just one use case that I think is very helpful for a lot of everyday Hermes agent users. So, we're going to break this

users. So, we're going to break this video down into six different sections.

You can go ahead and skip to the timestamps for whichever one is most relevant to you.

But, it's going to be quite a long one.

But, I promise you, I think you'll learn a lot from this video. So, starting off, let's talk about my old cron job setup.

So, it was a very simple setup. It's one

cron job for researching every news in the AI space over the past 48 hours and generate a report summarizing those news. It runs every 9:00 a.m. Hong Kong

news. It runs every 9:00 a.m. Hong Kong

time. And this was how the pipeline went. It's

went. It's It just spawned a single sub agent running 14 web searches sequentially,

wrote a report, then updated my landing page, which I have right here, this one, and then post the notifications on Discord to see if it if

the run was successful or it failed. So,

so far, it worked pretty well, right?

But, of course, there's limitations there. First and foremost, I didn't have

there. First and foremost, I didn't have this in mind when my Hermes agent would just spawn one sub agent. I told my

Hermes agent to spawn sub agents, not one. So, I overestimated the

one. So, I overestimated the intelligence of my Hermes agent. It went

ahead and spawned one sub agent. And it,

you know, everything was running sequentially. If one search failed, the

sequentially. If one search failed, the whole pipeline could stall, and that's a big problem. There's also no separation

big problem. There's also no separation of concerns. So, the same sub agent

of concerns. So, the same sub agent would do all of the searching, all of the writing, and all of the publishing, which doesn't make any sense. Like, if

you if it's one agent doing everything, then shouldn't it be my main Hermes agent to do it all? Cuz at least my main Hermes agent has its all of the bootstrap files, all of the memories,

the system prompts. Even if it's overloaded, it would be able to produce something that is, you know, decent.

But, no, you know, the sub agent is worse.

An overloaded sub agent is worse because it only has the agents.md file and the tools.md file to boot in the session to

tools.md file to boot in the session to do its work. It has no memories. It has

no system prompts. So, they're literally there just to get the job done. And they

don't care about the quality. That's the

That's the main separation between the sub agents and your main agent. So,

maybe it was my fault. Maybe I could have been more specific with the sub agents. But, yeah.

agents. But, yeah.

She just digressing here a bit.

Single sub agent doing everything. And

there's no verification layer. There's

no retry logic. So, if the cron job failed, it wouldn't try again. It would

just completely stop. And then there's a date bug. This is quite subtle, but it

date bug. This is quite subtle, but it is very critical. So, the shell date syntax in my prompt was passed literally

to the search queries instead of being executed. Meaning, their search queries

executed. Meaning, their search queries for the AI news would be literally this: date year month

day, right? So, it's hardcoded. They

day, right? So, it's hardcoded. They

It's It's supposed to be 2026, May 5th, which is today. But, no, they put this, right?

As a human, you'd think, you know, duh, right? But, no, sub agents can be

right? But, no, sub agents can be really, really dumb in these cases. So,

as a result, you can see that my report size had been declining since May 2nd.

The reports were getting shorter, right?

So, let's go back to the version one pipeline. This was without the Kanban.

pipeline. This was without the Kanban.

You can see there's very few news. Now,

I know, of course, there's not going to be a new model release, at least from the big companies, right? Like,

Anthropic, Open AI. They release a new model every month or every 2 months.

But, on Hugging Face, it's full of information of new models, new local models. Why was that not included,

models. Why was that not included, right? So, at least here, right? In the

right? So, at least here, right? In the

new pipeline with Kanban feature, version two, there's more sources.

But, I still don't see Hugging Face. So,

that's needs improvement there. But, at least it's able to search more thoroughly on AppCode GitHub developers.googleblog.

But, it has potential because there is feedback loop here. So, it's very obvious to see that in the old pipeline, there was no self-improvement loop, right? I needed parallel execution,

right? I needed parallel execution, verification, and specialist workers.

So, Hermes Kanban was the solution for that. The timing could not have been

that. The timing could not have been better, which brings us to section two.

How does Hermes Kanban actually work, right? So, basically, Hermes Kanban lets

right? So, basically, Hermes Kanban lets you create multi-stage pipelines with specialist profiles. Now, we covered

specialist profiles. Now, we covered this briefly in the first Kanban video, but it's a very important distinction here that the parent task and child tasks of the Kanban feature is, you

know, way ahead of your traditional orchestrator agent delegating task to sub agents. And this is the pipeline of

sub agents. And this is the pipeline of my AI news daily, which is the parent task.

It spawns four parallel search task, one for model release, tool release, agent frameworks, and trending workflows. And

there's also a fifth parallel search task, but I don't know why it's not shown here. It's just checking active

shown here. It's just checking active inputs. So, it doesn't have to be for

inputs. So, it doesn't have to be for the cron as well. If I need a news report on demand ad ad hoc on demand, then that is also in there as well. It's

a very flexible web search skill. It

doesn't have to be just for models, you know, more flexible, right? Then once

all of the five parallel search task complete, the verification task unblocks. The editor would filter the

unblocks. The editor would filter the duplicates. They would check the dates

duplicates. They would check the dates and then rank the news articles by importance. And then the write report

importance. And then the write report task unblocks. So, what I mean by

task unblocks. So, what I mean by unblock is when this parent task starts spawning at 9:00 a.m. Hong Kong time for my cron job to fire,

all of these Kanban workers would get spawned. But, remember in the Hermes

spawned. But, remember in the Hermes dashboard that we showed you, you can actually check if they're blocked, if they're in progress, if they're ready, or if they're done. Basically, if it's

not their turn yet, they would be blocked. So, in real time, it would look

blocked. So, in real time, it would look something like this, where the researcher would be in progress. But,

the editors would be blocked, okay? But,

once they're done, then the editors would be in progress, right? But, the

publishers would be blocked. And then

once they're done, the the publishers would be unblocked, okay? So, the write report task would get unblocked, then the publisher updates the HTML,

and then finally the Discord notification unblocks. So, I get a

notification unblocks. So, I get a notification posted on Discord. So, when

I wake up, I know that I get my news report every morning. So, the whole flow of the Kanban pipeline is the gateway

would auto dispatch those workers as the parent task becomes ready. And then the blocked tasks would wait for their

parents. Failed tasks would retry. And

parents. Failed tasks would retry. And

this is the most important part. You can

have even up to 100 runs. It will keep trying and keep trying and keep keep trying. And at a certain point, if it

trying. And at a certain point, if it knows that there it cannot try again, it will let you know, and then that's where it needs human input. Okay, so far, you've seen how much stuff we have to

get set up for Hermes agent to work correctly. And for the everyday Hermes

correctly. And for the everyday Hermes agent users, this is certainly not ideal. If you want to focus only on

ideal. If you want to focus only on building and not wasting time fixing stuff or configuring, then deploying Hermes agent on a Hostinger VPS is the

right fit for you. We want to thank Hostinger for sponsoring today's video.

Now, the standout feature of deploying Hermes agent on a Hostinger VPS is their ready-made Hermes agent Docker template, which handles the entire deployment

infrastructure automatically. So, if you

infrastructure automatically. So, if you really want to start building serious work with your Hermes agent, use my link hostinger.com/boxmining,

hostinger.com/boxmining, which will redirect you to this page, and we recommend the KVM 2 plan for running your Hermes agent comfortably.

So, click deploy, and once you're here, we recommend that you also get the daily auto backup to get your VPS data backed up every day because given Hermes

agent's self-improving nature, where it builds persistent skills, memories, and session history over time, the gap between weekly and daily backups is substantial. And choose a server

substantial. And choose a server location that is closest to you for the best performance. Use coupon code

best performance. Use coupon code boxmining to save 10% off of your annual plans. This applies to the 12-month and

plans. This applies to the 12-month and 24-months KVM plan. So, if you really want to get the most amount of value from your Hermes agent and start building amazing work, check out my link

hostinger.com/boxmining

hostinger.com/boxmining in the description and pin comments down below, and apply code boxmining for an exclusive discount for all Hermes agent KVM plans. All right, back to the video.

KVM plans. All right, back to the video.

Okay, so let's move on to section three of how I got it working. So, the first thing to do is to create the specialist profiles, configure the tool sets, and

then set up the API key. Now,

apparently, from what I did, you don't need to do these manually because I just went directly in Hermes TUI and told my

agent, the first prompt that I did, all right, was I told my agent, "Can you study and understand the official

documentation of the Kanban uh features?" So, here.

features?" So, here.

I copied this link, and I copied this link, send it to my Hermes agent. I tell him, "Can you study and understand this first?"

first?" And then, how can we apply the Kanban feature to improve our cron job? So, it

was able to uh understand what the Kanban is, and I think this is a very overlooked step. Is don't assume that

overlooked step. Is don't assume that your Hermes agent knows everything about Hermes agent. Feed them the document

Hermes agent. Feed them the document first, right? Even though they are

first, right? Even though they are Hermes agent, they might not know what a Kanban is, all right? Maybe they might do when when you ask them, "Can you research the Kanban feature?" They might

look for uh you know, they would do web searches from Reddit or, you know, another platform, but they might not check the official documentation. So,

it's a good habit here, whenever there's a new feature from Hermes agent, is to link, okay, the official documentation, send it to them, ask them to understand

it, and then ask, "How can that apply to our pipeline, right?"

So, I asked that, uh which it was able to spin up all of these, okay? The

profile creation, tool set configuration, and the API keys, okay?

And the API key is a bit tricky, okay?

Because when your agent, or even when you create the profile, the dot env file is completely empty. It is not auto-copied from your main agent's uh

dot env file. So, when it created all of these roles, uh it had asterisks in the profile.env.

There was no API key. So, what you need to do here is you need to remove any empty API key field from the config.yaml,

config.yaml, and then copy the real key from the main your main agent's dot env to each profile, okay? You don't Okay. So, if

profile, okay? You don't Okay. So, if

you get confused here, you don't need to do this manually, okay? I already gave you the documentation here. You just

copy this, and then you tell your agent in TUI, uh "Is this the correct step?" Okay? And

you can actually reuse the same API key.

You don't need to have 20, 30 different API keys.

Uh this is especially handy for people who have coding plans. So, we have the Kimi uh coding plan, and we just use the same uh API key for all of the roles.

Now, in terms of how much tokens they would use for the task, uh crank it uh 100 being, you know, most tokens used.

For research, of course, I want good quality, I would have 90 to 100. So,

whenever the it requires critical thinking or researching, then that would be the max. Uh but for editing, you can dial it down a bit because it's just about synthesizing the info gathered

from the researcher, which doesn't need much, you know, brainpower. Oops. Uh and

then publisher, you can even have this down to 20, 30, you know, etc. But this is, you know, what my Hermes agent thinks is best for these profiles. So,

after your Hermes agent would do this, the next step for you or your Hermes agent, okay? You don't need to do this

agent, okay? You don't need to do this manually, all right? Just we're going through the process here. Uh is your Hermes agent would create the parent

task, and from these parent tasks, your agent will determine how many uh children or Kanban workers would be needed to accomplish all of those tasks.

So, from my pipeline, it uh spawned nine children or nine Kanban workers. And

then, it would link the dependencies, started the gateway, and then start the test run. So, um the comparison of the

test run. So, um the comparison of the reports here between the old cron and the new cron, the new cron with, you know, with the

Kanban, as you as you just saw, it was way more structured with proper tables, categorization, and 48-hour uh filter verification. Now, let's talk

about section four, the limitations. So,

Kanban, all right, is powerful, it's great but uh specifically for cron jobs, a lot of stuff broke. And in hindsight, I

stuff broke. And in hindsight, I realized if I were to redo this again and try, I would not use the Kanban feature for the cron job. Uh but I was actually uh able

cron job. Uh but I was actually uh able to find a solution for this. So,

here's what happened. The biggest

problem here is the gateway exits. So,

what happens is the uh Hermes gateway start with dispatch the ready tasks, and then exits.

Meaning, if a task is waiting, okay, for its parent, and the parent completes after the gateway exits, then the block

task never gets dispatched, right? The

child just sits there forever waiting for the next move. And you can actually see, if you run Hermes Kanban list, uh the gateway is dead. It's it's ready,

but it never gets dispatched, so the pipeline stalled. So, my Hermes agent

pipeline stalled. So, my Hermes agent was able to fix this uh with a systemd service. That way, the gateway stays alive permanently, and

this is very, very effective for VPS cuz you're on 24/7. If you're running this on local, if you're going to try this fix on local, ooh, good luck with the bills.

[laughter] Now, the second problem here is during the test runs, uh my agent would keep creating new parent tasks without checking if one already

existed for that date. So, it would duplicate task creation. In fact, this is actually what happened uh today, right here.

So, today, May 6th, 2026, 9:10 a.m.,

it created the task, but it was labeled wrong. It's May 5th. And then, it

wrong. It's May 5th. And then, it created another task, but this time with the correct date, but it it's the wrong link. When you go here, it's not the

link. When you go here, it's not the right domain. It's not

right domain. It's not loki.boxmining.one.

loki.boxmining.one.

And I suspected this was because of the orphan tasks from the test runs that uh that I ran with my agent cuz I wanted to test it first, and then if everything is

okay, okay, then let's proceed with the cron the next day. But what what ended up happening was these uh test runs were actually orphan tasks. So, when the cron

fired today, multiple task sets for the same day all completed and sent Discord notifications. So, that's how I got the

notifications. So, that's how I got the two notifications today. Now, the third problem here, and it's actually the root uh problem that I don't know, maybe I

feel like in the future, the Hermes agent team will uh fix this, but there's just no synergy between Kanban and cron jobs. Because

cron jobs are designed to fire on a schedule right?

Because of that, they do not check the Kanban board before creating tasks. So,

without adding my own deduplication uh logic, every cron run creates a fresh task set regardless of what is already running or

completed. And the last problem is the

completed. And the last problem is the task accumulation. So, completed parent

task accumulation. So, completed parent tasks stay on the board forever unless archived. And after a week of daily

archived. And after a week of daily runs, I would end up having seven parent tasks and 63 children cluttering the

list. Damn, 63 children. All the with

list. Damn, 63 children. All the with all them child support, right? So, this

makes monitoring very hard and increasing the risk of accidental re-dispatch, which I anticipate in the next Hermes agent update, they will address that as well. Cuz remember, if

you go to the Hermes agent dashboard, you're actually not able to delete. I

tried that in the first video, I couldn't delete it. I could only block it, right? There has to be a delete

it, right? There has to be a delete button uh introduced soon. All right, so that's it for the setup. Um

if you are still confused from all of that, uh we can go through the lessons learned here, but let me give you a much more easier

feedback here. Is if you're going to try

feedback here. Is if you're going to try out the Hermes Kanban, just do it for um a non-cron job project, okay? So, you

what you do is you go to the workspace of your Hermes agent, and then just create a folder here, right? So, I

created a folder called AI News. This is

basically uh everything every files related to the AI News. That way it's, you know, cleaner, much organized approach. Then what you can do is go to

approach. Then what you can do is go to terminal and just go CD Claude, or you know, whatever the name of your workspace is. This might be different

workspace is. This might be different for you because Low-key was migrated from Open Claude way back when it was called Claude Bot, so the workspace name is still Claude.

Uh you know, just I whichever workspace directory you have, you just CD in there, and once you see it went to that project, that's when you can uh open Hermes.

All right, so your your um Hermes agent knows immediately all of the files in that project. So, it's a cleaner separation, and then you can build out the pipeline here using the

Kanban feature. And for the lessons

Kanban feature. And for the lessons learned here, I'm not going to go through all of them, but the thing here that I learned the most is there's no natural synergy between the Hermes

cron and Hermes Kanban cuz it needs the that deduplication logic for it to work properly. And as you can see, I had we

properly. And as you can see, I had we had four problems here that the has its own custom-made solution made by my uh Hermes agent, right? Four of them just

for this to work. And honestly, I don't know if I'll just keep this running until, you know, the Hermes agent team finds a way to integrate a delete button

for the Hermes tasks, then I think I would just stop this. But if they do have that delete task, then it's going to be a much cleaner approach moving forward. So, yeah, to end off this

forward. So, yeah, to end off this video, this is the current architecture that I have right now. Uh the cron job, okay? So, when whenever it's 9:00 a.m.

okay? So, when whenever it's 9:00 a.m.

Ho Ho Time, Hermes cron fires, and the first thing my agent does is it do it does a deduplication check. Wow, that's

a lot of alliteration.

All right, so it it uh searches for existing tasks to make sure it's not conflicting with the report today. Okay? Then great, there's

report today. Okay? Then great, there's no uh error there. Okay, now it spawns the Kanban parent task along with the nine children, uh or those Kanban

workers, okay? So, five tasks for

workers, okay? So, five tasks for researchers, two for editors, and two for publish. And then for for the

for publish. And then for for the gateway to work properly, it it uses the system MD service to keep it alive permanently, to make it persist permanently.

All right, [clears throat] so yeah, that's pretty much it for this video.

We're definitely going to do more guides on the Hermes Kanban. We'll consider

this a part two of the Kanban uh guides.

Uh I'm very excited to use because practically using uh communicating with your Hermes agent in TUI

and leveraging the Kanban feature is almost the same feel as talking to an agent in an IDE, right? Like like on Claude Code. But now that Claude is

Claude Code. But now that Claude is nerfed, right? We've been using Kilo

nerfed, right? We've been using Kilo Code, but honestly, this one is just as good as using uh Kimmy on Kilo Code as well. So, yeah, if you find this video

well. So, yeah, if you find this video helpful, smash up that like button, subscribe to the channel. My name is Ron. Shawn and out.

Ron. Shawn and out.