AgentCraft: Putting the Orc in Orchestration — Ido Salomon

Ido Salomon, creator of AgentCraft, MCI, and co-maintainer of MC Apps, presents a game-inspired multi-agent orchestration framework designed to address the fundamental bottleneck in scaling AI agents: the human operator. The core thesis is that running more agents in parallel does not automatically yield proportional productivity gains because humans become overwhelmed managing and babysitting many simultaneous agent sessions. The talk argues that the skills required to overcome this are not entirely new — they have been hiding in real-time strategy (RTS) gaming, where managing dozens of units in parallel is a practiced craft.

Salomon introduces AgentCraft through a live walkthrough. The system provides a spatial, game-map interface where each agent is a physical entity on-screen, and the file system is projected as a navigable map. Directories appear as zones and individual files appear as rooms, allowing operators to visually track which files each agent is actively modifying. This affords full lineage: who (which agent) changed what, and when. A heat map overlay visualizes potential file-system collisions between agents and can proactively surface conflicts before they occur — a direct import from RTS-style situational awareness.

The tool supports multiple input modalities — voice, text, and images — and can spawn coding agent sessions directly (e.g., Cursor, Claude Code, Codex, OpenHands). A mission-status side panel surfaces high-level summaries, while an RTS-style muscle-memory cycling mechanism allows rapid context-switching between agents awaiting human input, such as plan approval or clarification.

Beyond basic orchestration, Salomon addresses cognitive overload: operators have a finite number of concurrent ideas they can hold and act on. AgentCraft addresses this with a quest system where agents autonomously discover and propose tasks (refactoring, test generation, etc.) that operators can trigger with a single click. A campaign feature allows operators to describe a broad goal; a container spins up, agents decompose and plan the task internally, surface a plan for human review, and proceed — shifting human effort from continuous supervision to upfront planning and final review.

The campaign orchestrator pattern further enables cron-style autonomous pipelines: the demo shows a workflow that periodically scans Twitter for interesting ideas and automatically opens implementation PRs. Human review is then facilitated through review bundles — grouped PR views with diffs, task context, screenshots, and video evidence of agent actions — so operators can assess outcomes quickly without deep code review.

A standout feature is cross-user collaborative workspaces. Two operators (e.g., an engineer and a product designer) share a unified map where each can see the other's agents working in real time, observe file-level activity, and execute seamless handoffs — the designer's agents pass work directly to the engineer's agents. A shared chat channel accessible to both humans and agents provides soft coordination signals (e.g., an agent broadcasting that it is starting work on a file), enabling collision avoidance without hard locks.

Salomon concludes that the shift from writing code to directing agents does not require entirely foreign skills — gamers already practice them. The stated goal of AgentCraft is to raise the ceiling of human-agent collaboration through better visibility, greater agent autonomy, and richer human-to-agent and human-to-human collaboration patterns. The tool is available as a free experimental download with an active Discord community.