AgentCraft: Putting the Orc in Orchestration — Ido Salomon
TL;DR
The bottleneck in multi-agent systems is the human, not the agent count — Ido Salomon argues that spinning up 10, 20, or 100 agents is easy, but actually orchestrating “dozens of reckless employees” is the real problem engineers aren’t trained for.
AgentCraft borrows from RTS games to make agent orchestration feel native — the core idea is that skills gamers already have, like managing many units, cycling attention fast, and reading a map, can transfer directly into productive human-agent collaboration.
The standout UI metaphor is a live map of your filesystem — directories become map regions, files become rooms, and you can visually see which agent changed what, when, with lineage tracking and even a heat map to catch code collisions before they happen.
Salomon’s next step is reducing human babysitting by pushing agents toward autonomy — AgentCraft adds “quests” for agents to discover work on their own, and “campaigns” where agents decompose tasks inside containers so the human shifts from constant supervision to planning and review.
The endgame isn’t just one person with many agents — it’s shared human-agent workspaces — a product designer can run agents on their own machine, Salomon can see them, hand work off to “our agents,” and both humans and agents coordinate through shared chat and file awareness.
AgentCraft is positioned as an experimental free tool, not a finished answer — Salomon closes by inviting people to download it, join Discord, and help shape a system aimed at raising the ceiling on visibility, autonomy, and collaboration with agents.
The Breakdown
The pitch: agents are great, but humans choke on orchestration
Ido Salomon opens with a fun provocation: if one agent is amazing, why not use 10, 20, or 100? His answer is blunt — the problem isn’t spinning them up, it’s us. Engineers aren’t used to managing “dozens of reckless employees,” so the real missing skill is orchestration.
Why games, not enterprise software, hold the missing playbook
Instead of presenting orchestration as a brand-new discipline, Salomon says the skills have been hiding in plain sight: games. If you’ve ever managed lots of units in an RTS, AgentCraft is built on the bet that those instincts — attention management, rapid switching, spatial awareness — can be repurposed for software work.
AgentCraft turns coding agents into visible, physical actors
He demos AgentCraft as a kind of command center where a coding agent is no longer abstract — it’s a visible live session tied to tools like Cursor, Claude Code, Codex, or OpenClaw. You can spawn an agent, prompt it by voice, text, or image, and watch it start working while surrounding “buildings” in the UI expose skills, plugins, terminal access, and Git workflows.
The cleverest metaphor: your filesystem as a game map
The big visual idea is that the map is literally a projection of the file system: directories are places, files are rooms. That lets you see exactly where an agent is working, inspect the change list, and trace full lineage of who changed what and when. Salomon then layers on a heat map to visualize file collisions — and potentially prevent them before agents step on each other.
Visibility helps, but constant cycling still burns out the human
Borrowing again from RTS games, AgentCraft lets users use “muscle memory” to jump between agents that need approvals, answers, or plan reviews. But Salomon admits this only solves orchestration for short bursts. Eventually, the human still runs out of ideas and patience from all the babysitting.
Quests and campaigns: handing more initiative to the agents
His answer is to offload both ideation and supervision. Agents can generate “quests” like refactors and tests on their own, and broader “campaigns” let Salomon state a goal, spin up a container, and let agents decompose and execute the work there. The shift is key: less energy spent micromanaging execution, more spent reviewing outcomes.
From reviewing plans to reviewing many PRs with evidence
Once agents can run more independently, the bottleneck moves again — now there are just lots of PRs to inspect. AgentCraft responds with review bundles that show what changed, why it changed, and even visual proof like screenshots and videos, so Salomon can compare multiple attempts and pick the one that fits best instead of overthinking the initial plan.
The final move: shared workspaces for humans and agents together
Salomon’s favorite feature is collaborative workspaces where another teammate — like a product designer — can run agents on their own computer while he sees their activity and continues from where they left off. The collaboration is both explicit, through direct prompting, and ambient, through shared chat where agents can say things like “I’m starting to work on something” and humans can signal they’re touching it too. He ends by framing AgentCraft as experimental but free, inviting people to try it, join Discord, and help “raise the ceiling together.”