Speaker A
Many people talk about multi-agent workflows, but they're very difficult to actually pull off without breaking.
Learn how to build a durable, event-driven multi-agent workflow using Hermes Agent and a Kanban board for autonomous task coordination.
Key Takeaways
- A Kanban board can effectively coordinate multiple autonomous agents without conflicts or duplicated efforts.
- Event-driven task management enables self-healing and parallel processing in multi-agent workflows.
- Durability and auditability are critical features for reliable multi-agent systems.
- Hermes Agent profiles allow customization of agents with different models and roles within a workflow.
- Open sourcing the workflow template encourages community adoption and experimentation.
Summary
- The video demonstrates setting up a multi-agent workflow using Hermes Agent and a Kanban board to coordinate autonomous agents.
- The Kanban board acts as a single source of truth, preventing task conflicts and duplicate work among agents.
- Each task is represented as a card on the Kanban board, assigned to specific agents who claim, work on, and complete them.
- The dispatcher loop manages task assignment and agent spawning, enabling smooth, automated workflow progression.
- Tasks can wait for dependencies to complete before progressing, allowing parallel work and automatic triggering of subsequent tasks.
- The system is durable, surviving restarts and crashes, and is fully auditable with logs of claims, comments, and completions.
- The workflow example includes agents specialized in scouting, researching, analyzing, building, testing, and video production.
- The orchestrator agent acts as the central judge and driver, managing the overall workflow pipeline.
- The video creator plans to open source a generalized version of the workflow template for public use.
- Membership options are offered for early access and exclusive content to support further development and experiments.
Full Transcript — Download SRT & Markdown
Speaker A
In this video, I'm going to show you how I actually pulled this off where I had agents that did research, made judgments, built tools all autonomously with my human gate only being the final approval.
Speaker A
And I did this all with Hermes Agent and the Kanban board, so that you can have a whole fleet of agents working for you 24/7, doing research, building solutions, producing deliverables, making you money, operating your businesses, and much, much more. So watch till the end.
Speaker A
I'll show you how I pulled this off, and then I will open source my entire workflow template so you can try it out yourself.
Speaker A
Let's get started. And if you like this video, please consider supporting the channel by joining Team Garage, which will give you early access to videos, or Team Garage Max, where you will get exclusive videos each week, as well as
Speaker A
many other perks. I'd like to continue providing as much free, valuable content as possible, as well as better experiments on different hardware. So, these memberships will really make it possible.
Speaker A
So, one of the big features of the last few updates of Hermes Agent has been this Kanban, and you can find it if you open your web dashboard here.
Speaker A
It's under plugins, Kanban. And this is what it's going to look like. Mine looks a little bit different, but generally it's going to have these different cards that you will see.
Speaker A
So, this is a very exciting development because it allows a lot more multi-agent workflows that a lot of people are interested in setting up when they are using agents.
Speaker A
And the first version had some issues, and with every new update, they've been kind of working on it, and I'm sure in future updates, they'll continue to kind of refine it. But I've been using it now to set up this workflow and
Speaker A
it works fine. So, whatever issues were in the earlier versions seem to be ironed out.
Speaker A
So, in today's video, I'm going to give you a breakdown of what the Kanban is and show you a live example of me actually using it for a workflow that I designed myself and that I use myself.
Speaker A
And at the end, I will open source the workflow, some of the key structures, and scripts needed for a generalized version, not my specific workflow, but one that can do work for you using the Kanban and using multiple agents. This
Speaker A
video, I'm going to cover the Kanban board itself, how a fleet of agents coordinates without chaos.
Speaker A
The workflow that I set up just as a demonstration, right? And I'll show you how this actually works, how I set this up, and what it looks like in the Kanban here in the dashboard. You can
Speaker A
already see some work has been done. And then at the end, I'll show you the finished deliverables.
Speaker A
[snorts] So, first about the board. So, here's the hard thing about multi-agent systems that a lot of people experience. It isn't setting up the agents themselves, it's getting them to work together on one job without stepping on each other.
Speaker A
And this is really what the Kanban board is trying to solve. There have been many, many attempts at multi-agent systems and they've all had some success to a certain degree, but they all eventually run on the same issue, which is how do you get them to
Speaker A
work together without running into any conflicts with the work? So, without a Kanban board, you have agents racing around and doing duplicate work, wasting your tokens, wasting your money, right? No one has a shared memory of the progress and one crash loses
Speaker A
everything. With the board, every unit of work is a card, like you saw, these different cards here.
Speaker A
The agents claim it, work on it, and then hand it off. So, state lives on the desk and it survives a restart. The board is a single source of truth.
Speaker A
So, there's no chatting between the agents. There's no message queues or anything like that. Just the board that they all read and then write.
Speaker A
So, this is the design of the board. I showed you the real thing right here.
Speaker A
These different cards. But this can kind of show you the concept. And these are all the different sub-agents. Builder agent has this to do to build a prototype. These two cards are ready, the researchers.
Speaker A
This is the current one that's running, make a slide deck from the video producer agent. And these are the ones that are done.
Speaker A
So, this is one SQLite file. That's the entire coordination layer. And it's the bus and audit log all in one.
Speaker A
So, here's an example of a task. And they keep it very simple. It's just the title, what the task is, who's in charge of it, and what its status is.
Speaker A
So, the assignee is for routing. A card assigned to the builder runs on the builder agent.
Speaker A
So, this one is assigned to video producer, so it runs on that agent. And each of these agents are really just profiles inside of Hermes Agent.
Speaker A
We'll get into the profiles a little bit more, but it allows you to give different agents different models and different tasks.
Speaker A
So, how does a card become a running agent? It's one loop and it's called the dispatcher.
Speaker A
So, you can see this diagram from top to bottom. The board has a ready task.
Speaker A
The dispatcher claims it, so two never grab the same task, and it spawns the assigned agent in its own clean workspace.
Speaker A
The agent then does whatever the work it's been tasked and then marks the card as done.
Speaker A
And then it loops, so every tick continuously. And there's some really clever architecture to how this Kanban board works.
Speaker A
Tasks can wait for each other automatically. So, you see these different cards. The card will sit in to-do until its parents are finished. Then it promotes itself to ready.
Speaker A
So, this one task fans out to three research tasks. You verify the sources, pull prior context, audit existing solutions, and this all happens in parallel.
Speaker A
And I hang out route task on all three of these. So, the moment the last one finishes, the route fires itself.
Speaker A
So, you have three agents working at once and the next step triggers itself. So, there's no polling or no babysitting, no glue code holding it all together. It's all just done automatically.
Speaker A
So, why does this win as a multi-agent system? Why go through all this trouble?
Speaker A
There are five things. One, it's more durable. It survives restarts and crashes. The board is on the desk. Everything here can be recovered.
Speaker A
Parallel, you have many agents working at once all coordinated by one board. Event-driven, the work flows itself through the graph.
Speaker A
There's no polling loops or anything like that. It's self-healing. A dead task gets reclaimed and then respawned.
Speaker A
And it's auditable. Every claim, comment, and completion is fully logged, so you can go back and see exactly what happened and if there were any kind of issues.
Speaker A
Okay, so now I'm going to show you how this actually works with this workflow I designed myself. And I designed this in a combination of Claude and Obelus 4.8 as well as with Hermes Agent itself using GPT 5.5. So, this was the problem that I wanted to solve: find real pain points AI agent users hit, then do something about it.
Speaker A
This is the process. Detect: we have scout service complaints from across the web.
Speaker A
Validate: score each against a rubric, drop the weak ones automatically. Route: build a fix or make a video explaining a solution, and then ship after exactly one human approval.
Speaker A
And this is my fleet of agents that is going to make this workflow possible.
Speaker A
Every agent here is just a profile on one machine. The two scouts on top here, one is doing research on X, and this one specifically has a Grok model. All the others are GPT-5.5 right now, but it can be
Speaker A
whichever model you prefer. There's also the web research scout, which checks Reddit, YouTube, the web.
Speaker A
The orchestrator in the middle, who's kind of the central pipeline, this orchestrator is the judge and the driver. It runs everything.
Speaker A
Then you got these workers. You have the researcher, analyst, builder, tester, and video producer.
Speaker A
And these are all ones.
Speaker A
And these are all ones really doing the work. So, that's the main pipeline. You have scouts creating doing research and creating reports, orchestrator deciding what to do after that, and then all these other sub-agents doing the work based on that judgment.
Speaker A
So, here's the full pipeline top to bottom. Got this nice little diagram you could see. Starts with the scouts, and this is done hourly or maybe every 2 hours, depending.
Speaker A
Uh but this is a research on X and web, uh Reddit, etc. Then the orchestrator will intake those reports, and then it has to decide. It has to deduplicate. Obviously, you're going to run into a lot of the similar issues.
Speaker A
So, it has to deduplicate anything it's already seen before. And then we created a rubric, and it was scoring based on how often the issue had been seen, if it was something that was actually fixable, if it was something
Speaker A
temporary. We had a few different uh ways to measure this, but anything under a 65 out of 100, it was shelved and just not touched.
Speaker A
So, after anything that passes this gate, goes to researcher agent, and there there's three of them like I showed you before, and they work in parallel to verify the issue, figure out the context and existing solutions. Then the orchestrator, once again, gets this
Speaker A
report and decides to be build this, is this a tool that I can build?
Speaker A
Uh should I explain it as a video idea? So people can overcome whatever issues they're facing or shelve it, is it just not usable for whatever reason.
Speaker A
So then if it's a build idea, we have the analyst agent that synthesizes the information and then creates the idea in more detail.
Speaker A
If it's the video idea, we have a video producer agent that does research on what it could do and creates an outline for me. Then, and this is very important, this is the one human gate.
Speaker A
Everything else here happens autonomously. I'm not touching any of this. This human gate is very important and this comes out in Telegram and this is what it looks like. I'll show you in our live demo, but this is
Speaker A
kind of what it shows. Four proposals are awaiting your approval. And it has each of the proposals. This case they're all video.
Speaker A
And you have to approve them, shelve them or just modify them if you want to change the plan.
Speaker A
But this is very important to keep. You could get rid of this, but I would not cuz I just don't want my agents wasting all my tokens and wasting a lot of money building random stuff. So you want to at
Speaker A
least approve this decision. Because even after all of this, they're still going to have ideas and proposals that uh just don't make sense.
Speaker A
So after you approve it, obviously if you shelve it, it goes away. Uh if you approve it, then it goes into a builder agent if it's a tool that I want to build and then a tester agent that tests
Speaker A
whatever tool it is or skill, whatever I'm happening to make here. And the deliverable for this is usually either a script or skill.
Speaker A
Um and for the video, it takes the concept, builds out a presentation slides and a script speaker notes for me to work on and then delivers it uh in the same thing, the same Telegram channel.
Speaker A
Either the final code that it built here or the slides and speaker notes. So, that's the whole workflow and I have this actually working.
Speaker A
Um whenever you get this, you usually get a decent product, but you still need to work on it, obviously, uh to mold it until it's actually ready for production.
Speaker A
But, it at least gives you some kind of pretty decent starting point. Um and this is the judging rubric. Just want to show you this more in detail uh cuz this is important as well. This is the one we designed.
Speaker A
And it's based on frequency of the issue, the pain intensity, how serious is this issue, is it solvable or explainable, uh what's the solution gap, and what's the strategic fit in terms of my channel and the the dev
Speaker A
work that I do do. So, the bar for actually shipping and moving on to the next phase is 65. This is still still tweaking this a little bit, but I have found so far that this gets you at a good balance.
Speaker A
And then once again, the orchestrator decides either to build it or try to make a video of it or just forget about it.
Speaker A
Uh so, there is one human gate, like I talked about before, and this is the only place you're in the loop.
Speaker A
Okay, so I'm going to run this for real so you can actually see it, not just me talking about it. And you'll see how it works in the Kanban. You'll see the cards popping up.
Speaker A
And I will show you the full process start to finish. This is usually on a Chrome schedule, but I'm going to run it here so you can see everything work out.
Speaker A
Okay, to get this started, we're going to have to do this orchestrator gateway run cuz we need the gateway to be running or the orchestrator is going to send us Telegram messages.
Speaker A
And you can see it's got a warnings and stuff, but um this stuff's fine.
Speaker A
And you might want to do this in like a tmux type of shell cuz this is going to be a long-running ongoing gateway run. So, you want to keep this alive.
Speaker A
Okay, so I set this up as a cron, but I want to show you what the command would look like from the CLI cuz we're going to run it.
Speaker A
Force it to run right now. So, this is X research research agent. And it's using the pain point scout skill.
Speaker A
And the prompt is run one X pain point sweep now following the pain point scout X skill exactly.
Speaker A
So, this is going to do the research run. And we will start to see um the Kanban board populate once this is is through. So, I'm also going to do this for the web scout as well.
Speaker A
So, those are doing research and that'll take a little time. They are prompted to do pretty thorough research.
Speaker A
And then once they have completed their reports, they're going to send them in and you're going to start to see the Kanban board populate.
Speaker A
You see the scout agents are starting up. You see the sweep is complete. So, we should see task on the Kanban board and we do.
Speaker A
See this was claimed by a worker. The orchestrator has taken this report from the X uh the X agent.
Speaker A
And it's starting to work. The web agent is still working cuz it has more sources to kind of check over.
Speaker A
So, you can see this is the block itself. I didn't find any qualifying pain points in this window.
Speaker A
The X one is a little bit more narrow. We might be able to see more uh from the web scout.
Speaker A
Hey, you see the web scout is is ready. Just just came in and this one actually does have a couple of candidates for issues.
Speaker A
So, we'll see how the orchestrator judges them. It says six candidates and you can see them right here.
Speaker A
And the nice thing like I said is this is all orderable. Like it has all these reports exist as markdown files.
Speaker A
So if I wanted to go back and take a look at them myself, they're all here.
Speaker A
So you can see the one extra report is it went into the done section down here because there's no there's no candidates, there's nothing to do.
Speaker A
Now the web research report went from ready to in progress. So that's what I'm talking about with everything having a proper order to it.
Speaker A
Nothing stumbling over itself. So let's see what the orchestrator does with this one. And there's a worker log down here so you can see what it's actually doing.
Speaker A
You can see now the orchestrator is looking at these different issues. Specifically, these are all all six, right? Yeah, there were six candidate issues.
Speaker A
So it's going to judge each of these to see if it's something that it can actually use. You can see it's working down here on it.
Speaker A
And it's scoring it and doing the the deduplication. I originally set this up to be hourly, but it probably doesn't need to be hourly. I think that's probably overkill.
Speaker A
Maybe a couple times a day. You can see it's panning out even more. Look at all this stuff in progress. The researcher sub agents have been launched. Can you see this? Can you see this? Can you see this?
Speaker A
Here, made it a little bit bigger. The researcher agents have been out like I said there's three of them per per issue. So source verifyer verifying the source issues.
Speaker A
Researching the context around it and any existing solutions. So you have each of these now in work.
Speaker A
So, these are all researcher agents. 18, we have 18 workers all working together in parallel without tripping on each other.
Speaker A
And that's really the beauty of this system. And you can see the previous tasks are all in the done uh card or group.
Speaker A
So, the orchestrator finished that work and now it's up to the researcher. You could see the orchestrator now has several to-dos as well.
Speaker A
And this is routing these issues uh to me in the Telegram once these researchers are are finished.
Speaker A
You can see they're starting to finish their work. So, once all three of the researchers per issue are done, I it'll go over here to the to-do.
Speaker A
And once everything is done, it'll send me the message in Telegram. So, you could see in this to-do a little bit more specifically about what it's doing. When all the parent research lines are done, which are the three that I showed before,
Speaker A
so this will make the judgment, right? Is this something we build? Uh is it good for a video?
Speaker A
Or do we shelve it? And then it'll move on to this ready. Now, there were two ready and I guess those were they finished.
Speaker A
You see it's moving along nicely. So, after the to-do, then it goes to ready.
Speaker A
And then the agent orchestrator, I should say, orchestrator agent makes the judgment. You can see more more tasks are in the done pile here.
Speaker A
Um so, you can actually see this by agent here as we have a couple different agents in process.
Speaker A
Um you could see the researchers are still working on the research down to six tasks. And the orchestrator up here has two tasks of its own.
Speaker A
So, a lot going on here, but this is all happening without me touching anything.
Speaker A
Once I have this set up, it all happens automatically. It's all automated. So, this is the one that's in progress right now. It's reading the reports and judging which path we're going to take here.
Speaker A
You see this one moved over to ready. And this is the video producer. So, I guess one of them did get through this issue Cloud code sub agents fail when many MCP tools are configured.
Speaker A
So, that would be a good. So, the orchestrator decided this would be a good for a video.
Speaker A
So, now the video researcher is researching the current existing solutions and then creating a outline for me to approve.
Speaker A
We should be getting Telegram messages soon. You see they're moving in. With ready, the analyst is synthesizing the problem.
Speaker A
Um so, this is actually this got to a build path. So, this is actually considering building a tool or skill.
Speaker A
Um and this is for Codex uh VS Code extension ignores config uh tunnel changes.
Speaker A
So, now it's analyzing this issue and trying to come up with a solution. So, we should be able to see both video proposals and a build proposal.
Speaker A
And the nice thing is this is all very visual. Like, you can very clearly see what's happening even though I'm not doing anything. You can see analyst is in progress, orchestrator researcher video producer. We now got four uh
Speaker A
four agents all working in parallel all together. Everything's in progress. They're all working together, but they're working on different problems, so they don't step over each other.
Speaker A
So, going back to this diagram, we're now in this phase, right? Next one, they'll be sending it to me to decide if they want to go through with it.
Speaker A
Okay, you can see some of the video outlines are ready. And then it'll be assigned uh to the orchestrator to send it to me in Telegram when everything is ready.
Speaker A
Okay, we can see it starting to get proposals here. This is in Telegram uh with my Hermes agent Bulls here, you remember from Agent Trench's trial.
Speaker A
Uh Bulls is my Hermes agent on this device. But the first build proposal is building a tool uh and the pain point is the Codex VS Code extension users report safety critical mismatch uh between their intended config camel
Speaker A
approval settings and the IDE extension's behavior. So it did Here's the sources in GitHub.
Speaker A
So why build it? And giving it the reasons, it seems to be a frequent issue. Existing solutions are broken.
Speaker A
Uh bounded build scope. So what they would want to build is a Python script under 500 lines.
Speaker A
Um which is what we want. We don't want like full, you know, API thing.
Speaker A
So this is the proposed solution. Build a local Codex VS Code safety config verifier set CLI.
Speaker A
Um it would inspect likely codec config locations across Linux WSL and Windows home paths.
Speaker A
Uh should explicitly avoid claiming the CLI enforcement through VS Code extension enforcement. Okay. So So this is the build proposal and it has a pretty good outline, you can see, uh written by the agent.
Speaker A
So this is how you would reply to it, how I've designed it. You just simply do approve, you just say approve this.
Speaker A
If you don't want to do it, you can just do shelf. You can add a reason if you want to check the logs later, or you could just modify the plan.
Speaker A
Um Let's Let's build it just for a demo here. So, I'll do approve so you can see how it builds.
Speaker A
So, you could see while we got that report, it's still working on the issues. Oh, you could see in real time the builder started working on that prototype build.
Speaker A
So, this is connected with Telegram, but only through the orchestrator and only through the one gateway here.
Speaker A
So, post approval build chain prototype build builder ready, tester, and then final report and the final deliverable, which I think in this case it's going to be a script, a CLI um that I can work I can test out myself. I do have Codex.
Speaker A
I don't think I have the VS Codex extension, but I can probably test this out myself. Okay, we got more proposals here in Telegram. The cloud code sub agents fail when many MCP tools are configured. So, this is a video proposal. The existing
Speaker A
solution, there's not one simple guide, but there are usable pieces spread across Anthropic's docs.
Speaker A
So, why a video would help specific friction, users see many silent-looking, you know, sub agent exits.
Speaker A
Instead of understanding the sub agent was born with a giant MCP tool manual in its prompt.
Speaker A
So, this is a proposed video. So, it would be a guide for people. Um Take a look at this.
Speaker A
These are all the sources. You see a lot of people having issues with this.
Speaker A
Might be a good video. So, just for an example, you could see we already did the build one. Let's approve this one as well.
Speaker A
Uh so, approve this so you can see both of the pipelines. So, it's approving this as a video. So, it's going to go back and research.
Speaker A
So, this is the next approval, right? Yeah, cloud code sub agents, silent leave fail rights and MCP calls.
Speaker A
Okay, so you can see it's approved here. And the video producer is starting to work on it.
Speaker A
And it's based on the research, it's going to go once more over the research.
Speaker A
So, now we're seeing the human gate has gone through. I approved two of them.
Speaker A
Right now, it's building and we have another uh pipeline creating slides and a script based on the research.
Speaker A
This is during the testing phase, but I just wanted to show you this cuz I thought it was interesting. So, one of the features about this set up with the orchestrator agent is that it has this kind of self-healing
Speaker A
factor to it and it's one of the reasons why you want a model uh in the loop, several models actually.
Speaker A
Because in this testing phase, I have this ironed out by now. Uh but it had delivered the slides for the video concept. But the slides have been you can see the problem here.
Speaker A
Um the approved post gate slides and script were written to scratch task work places that are no longer present.
Speaker A
And they're just like temporary work places. So, the final delivery cannot reference real files. These are just temporary.
Speaker A
So, it realized this it was an issue. And I didn't tell it this, it did this autonomously. It had this self-healing function where it regenerated the video slide deck again to a persistent output directory.
Speaker A
So, this This a bug that I will have fixed by the time you look at this actual code base and in my primary example, but I just wanted to show you this this is during testing uh cuz this is built in.
Speaker A
This specific issue has been solved, but if you run into any issues with the deliverables, um it has this kind of recovery function recovery task uh that is is very useful.
Speaker A
Okay, so let's let those build out. And you see the outline of the video they're describing here and based on this research, uh first lever is verify modern tool search, second stop giving every sub agent every tool. So there's allow list
Speaker A
for custom sub agents. So this is something where a solution exists, it's just not always clear. So having a video um would make it more clear for people.
Speaker A
And this is why I designed this pipeline because I wouldn't know this problem existed. It's not something I've been encountering myself, but clearly a lot of people have been dealing with it.
Speaker A
So as a content creator and trying to educate, having seen what people are struggling with is a really good source to try to make videos and content that's really helpful for people.
Speaker A
So if there's a proposal you don't want, you just do shelf here and you can add a reason like I said.
Speaker A
And this will shelf. We're only going to do the two uh the two uh builds that I showed you, the one build in one of the video.
Speaker A
And you see this was shelved and that'll take it off the Kanban and put it into done.
Speaker A
Okay, so the builder this is in blocked, but it's not really blocked. It just needs my input. You can see it actually did under here uh it built the CLI, the config verifier CLI, plus a readme report template. All
Speaker A
the tests passed and the artifacts are on persistent workspace. So it's in blocked, but it just needs me to double-check it.
Speaker A
Uh but it did build the, I guess this is a a CLI tool. So, let's check it out.
Speaker A
So, it wrote out a nice read me for the tool here. So, with a quick start. So, I could very easily like open source this.
Speaker A
So, that'd probably be the next step. Yeah, and it's just a Python script here.
Speaker A
So, this auto discovers likely config paths for uh Codex VX code configs. Okay, I ran the script and this is uh it creates only a report. It doesn't modify anything, which is good. Uh, but it basically tries to inspect candidate
Speaker A
config.toml paths, redacts likely secrets, and records policy values, uh user-intended Codex to use. So, this is the result. I don't have any configs uh for this specific extension, but it did find the paths, right?
Speaker A
I found a couple paths here. So, it's not going to completely solve the issue, but it's at least a tool you can use to uh start to address it.
Speaker A
So, like actually building tools, and this one was a little bit complicated, but uh it's not quite easy to one-shot it, but this at least gives you a good starting point because all the research into the issue has been done, and it you have at least
Speaker A
have a a template or a basic foundation for the tool that you're trying to write.
Speaker A
So, we verified that the test works, so then you would just go in here and unblock this.
Speaker A
And you see it went into ready. So, we could see everything else finished working.
Speaker A
Um the tester still needs to test. And then we have a final report for that. I think the the video one finished while we were messing around with that. Without looking anywhere, the nice thing about Kanban is I could just look here in the
Speaker A
done, and I could see this card, and it is uh finished. So, let's see what it came up with for this is this is a video task.
Speaker A
Here we go. It created the slides. These aren't as beautiful as my my usual slides. So, I guess that this is just a a starting off point, but this is based on all the research.
Speaker A
Our sub-agent did nothing or did it die before birth? So, this is kind of explaining a problem where a solution does exist, but it's not always clear what it is.
Speaker A
So, like I said, this is not something I would just put out without reviewing it.
Speaker A
I would have to look closely at the research it did, uh look at the slide, certainly improve the visuals a little bit. Well, it doesn't look that bad.
Speaker A
And then decide how to tweak it from here, but at least gives me a really great um starting point.
Speaker A
And you can see the final step there is it in the Telegram. Just in case I'm out and about, I can't look at the Kanban itself. It gave me the final video deliverable message here. Telling me deliverables actually
Speaker A
are. Gives me a fact sheet, final handoff, markdown file, the slides, a script to go with it.
Speaker A
And there you go. So, you could do this all just by those one command, basically the one approval, and you would have this really nice, uh foundation.
Speaker A
And you can see we got our two deliverables. Uh see all the tasks we're up to 97 tasks done in that one run.
Speaker A
So, this is an autonomous system and it all starts based on those two scout researchers, who you can set up on a cron job. Like I said, I would probably do this maybe daily, maybe two times a day tops. Don't know
Speaker A
how often these kind of pain points come up for people. Feels like a lot, but maybe not so much.
Speaker A
Uh but this is an entirely autonomous setup, except for the one human gate. So, that's going to be it. I hope, uh you enjoyed seeing this kind of setup.
Speaker A
There's a lot of different ways to design this, connect it to your own workflows so that you can have researchers and you have agents building stuff without any kind of breakdowns, any kind of duplication or agents working over
Speaker A
each other, stumbling over each other, breaking stuff. You saw everything ran start to finish without anything breaking.
Speaker A
And if something does break, you saw that self-healing element to itself that it automatically noticed there was an issue and then fixed it, which is pretty pretty incredible.
Speaker A
And if you want to try this out yourself, I have open-sourced this workflow. I made it a generalized version so that you can adapt it to whatever purpose you have. It's under Tombi Studio Hermes multi-agent workflow.
Speaker A
So you'll have to work on this with your agent to get it to exactly what you want, but the skeleton is all here and it does the same basic path, right?
Speaker A
Where you do some sort of research, you do some type of scoring, and then you take that research and decide there's a couple different paths about preparing to do some type of action. And there's a human gate in the middle to
Speaker A
actually approve the action. So that's the idea, the same workflow, but you can just adapt it to whatever purpose you have.
Speaker A
So that's going to be it for this video. Please leave a comment. Let me know what kind of multi-agent setups you have, what kind of purpose or workflows that you're working on.
Speaker A
I know these are very interesting, but can be very difficult, very brittle. So we're always trying to find ways to make them more durable.
Speaker A
And if you like this video, please leave a like. Please subscribe. I'll be doing a lot more work in this kind of multi-agent setup agentic workflow space cuz I think there's a lot of room here. But that's going to be it
Speaker A
for this video. This one ran a bit longer than my usual videos, but I think it was worth it. And I will see you in the next one. Thank you for watching.
Topics:multi-agent workflowHermes AgentKanban boardautonomous agentstask coordinationevent-driven systemagent orchestrationworkflow automationAI agentsopen source workflow
