AgentCraft: Putting the Orc in Agent Orchestration

Hey everyone, welcome to AgentCraft, putting the Orc in Agent Orchestration. By the end of this session, you'll be able to call your parents and hit them with an I told you so. I'm Iddo Salomon, I'm an AI lead in Mand8's CEO office. In my spare time I am the creator of MCPUI and the co-creator and maintainer of MCPApps in the MCP steering committee.

Recently I created AgentCraft, which is why we're all here. Now, I don't know you personally, but if you're here, you've definitely used an AI agent. I bet that felt pretty magical. You open a terminal, you type something in, and it quickly produced work that would have taken you ages to do on your own.

So if one agent is so magical, why not we scale up and have 100 agents and get 100 times the returns? Well, the sad thing is that it's simply not that trivial. If we take this real-life example of scaling up to 25 agents, you'll notice a few things. One, that it's absolutely glorious. But the second thing is that it's not really scalable. It basically all falls down on one thing. Cognitive overload.

It's difficult enough to juggle different thoughts in your head while you're focused. But when managing a horde of different agents, you quickly lose track of what's going on. You context switch around just to get some visibility into what's happening, but your mind is focused on managing the agent instead of shipping. The conventional solution to this issue is, of course, an agent orchestrator. But what makes a great agent orchestrator? Well, there are a few different mechanics. The first and most obvious one is our ability to spawn different tasks. The second one is our ability to monitor those tasks. What's going on, what's stuck, what needs our attention. The third one is real-time management. We need to be able to react to those changes and steer our agents in the direction that we're looking for.

If you grew up playing video games, this might start sounding familiar. These three mechanisms are exactly what real-time strategy games were perfected for. In essence, RTS games are glorified task management simulators, and it's the simulators we've been playing since we were kids. We've built units, we've constructed buildings, we've protected our borders, and this is precisely why I created AgentCraft.

There's no way to manage agents like an RTS game. Let's stop talking and see something cool. This is AgentCraft. It's actually running on AgentCraft, so we're using AgentCraft to build AgentCraft, which is awesome. If we take a look around, we have this nice Warcraft 3 themed map. We have heroes here that are actually agents. We have these peons here that are actually subagents, and we have these nice buildings. Like any good RTS, these buildings are there to divide the different functionality.

For example, if we take this Warfactory, it equips our agents with capabilities, so it has access to the plugins and the skills, and we can enable, disable, do whatever we want. We also have the blacksmith, so we can install skills from a marketplace, from skills.sh, with a click of a button. Because AgentCraft is also where we work, then we are able to use it to actually craft new skills to improve our workflow going forward. So now that we kind of know what's going on with the interface, let's start working through the parts that make a Neo Crusader good.

The first one was just being able to spin up tasks easily. Right here I have, in the town hall, the ability to launch new agents. You just click, and I have a new cloud code agent running, and I can interact with it with this nice side panel, which is fully featured. It has slash commands. It has models and modes, and everything that you're accustomed to. And I have a feature that I wanted for a while, add multiplayer support, by the time that you watch this, it will already be implemented. And there are a few other stuff that I want to do. I can't think of all of them, because there's a lot of code going on here.

Luckily, I do get this nice suggestion list from AgentCraft, this quest board, and I just want to fix body validation, I want to add unit tests, I can just click Accept Quest, Accept Quest, and these new instances of cloud code are spinning up and doing it for me, and I can just monitor what's going on. So really the second part, after we saw that we can launch agents, and we can launch them in work trees, or main work tree, whatever we want, we just need to monitor. So we have this, it's kind of chaotic, if you look at the map, but like all RTS games, it kind of takes some getting used to, it's all very deliberate.

So if we look here at the nice widget on the side, we can see that there are a few heroes that are working, one is idle, one is actually awaiting input. Just by scanning, we can see these different question marks for heroes that need us, an exclamation mark for those that are idle, and those that are working or visibly working on something. We also have these sound effects so we know when something happens in the lifecycle. We also get notifications if the window is minimized. And when we take a step back, it's not only the agents that are doing stuff, we can also see that we have these nice rocks, kind of spawning on the map.

These represent files in your repository. So the entire repository is projected onto the terrain, so you have here the test directory and the docs and source, whatever. And this is a nice way for you to visually see what the agents know, and where they are. So if we toggle Fog of War, which is a feature that I had to add, you can actually see how they explore as they go. And with Heatmap, you can actually start seeing patterns and collisions. It's something that Agents Graph tracks, they track collisions and conflicts and solve them dynamically, so you don't always need a work tree to work and merge code.

If we turn off Fog of War, let's say we want to know what is the next thing that I need to focus on. I can just click the spacebar, and boom, I go to the first thing that needs my attention, and I need to say what supporting codex TLI means, which is a valid question. I can click space again and go to the next one with the primary RDN for the README, all developers. What does multiplayer mean, cross machine? I can manage in real-time, command heroes, stop them, simplify tasks, fork heroes, and hand over tasks. Agent Craft unlocks capabilities in a few clicks, facilitating code inspection, selection, and merging.

When delivering work, I can inspect individual agent contributions or view all agents' work to decide on merging. I can add, suggest a message, commit, and complete tasks efficiently. Agent Craft enables quick reactions to changes, unblocking heroes promptly and focusing on essential tasks. It unburdens users by offering quests to improve projects, summoning heroes instantly. The tool provides an enjoyable experience by simplifying complex tasks and offering a wide range of functionalities.

And the cool thing here is that we are really unlocking a wide range of capabilities in just a few clicks. So if we go and say that we want to actually deliver something that my agents worked on, I can see what this agent did and I can commit just that part or inspect just that part. Or I can see what all the agents did and decide what I want to merge. So I can just add them and even suggest a message, and commit and get my work across. So that was a blast. What did Agent Craft actually give us? So the first thing is that we're able to react to changes with muscle memory, I could just click space and get to the next agent that needed my help. And I could just answer by clicking a button and sending on its way. I was able to unblock three different heroes in five seconds, which is always good. And I could focus on what mattered. I didn't just focus on the management aspect. Agent Craft actually unburdened me by giving me a bunch of different quests that I can work on to reach my ultimate goal, which is just to improve the project. And I could summon these heroes just by clicking on it.

So I wasn't burdened by my own limitations as a human. And lastly, it's just fun. It's nice to be able to pick up scrolls as items, it's nice to have fog of war, it's nice to hear your heroes say prompt prompt when you ask them to do something. So it's kind of a whimsy that really makes me have that just one more hour before bed kind of feeling when I code. The way I look at it is, we finally reached a point where the skills we have as modern humans is crossing from leisure to productivity. Extreme multitasking is no longer for casual consumption or gaming. It's the way we do work. And it's more work than we've ever done before. Agent Craft is that bridge between those two worlds. And it's just the beginning for Agent Craft, there's so much more coming, there's multiplayer, there's disruptive features that will truly turn us into the best generation of builders. So thank you.

Please go to getagentcraft.com. This is the QR code. Hit me up on Discord or X or LinkedIn, I want to hear from you. So let's make work fun. Thank you.