Descriptions:
Ido Salomon, creator of AgentCraft and maintainer of MCP Apps, presents a live demo of his multi-agent orchestration tool at an AI Engineer conference in London. The central thesis is that managing dozens of parallel AI coding agents is less a technical problem than a human interface problem — and that real-time strategy games, not traditional dev tooling, offer the best design patterns for solving it.
AgentCraft visualizes the file system as a navigable map where each directory and file is a physical location, and each active coding agent (Cursor, Claude Code, Codex, or others) appears as an entity moving through that space. This lets engineers see which files are being edited, by which agent, and when — with full lineage tracking and a collision heatmap to detect and prevent conflicts before they happen. RTS-style muscle-memory shortcuts allow rapid cycling between agents that need human input, reducing the overhead of babysitting parallel sessions.
The demo progresses from basic agent spawning through increasingly autonomous modes: agents that self-assign tasks from a quest list, a campaign orchestrator that decomposes and runs multi-step work inside isolated containers, and a scheduled mode that autonomously scans sources like Twitter for ideas and opens PRs. Salomon frames the endgame as shifting human effort from constant supervision toward periodic planning and review — a workflow he argues is already practical with current models when the orchestration layer is designed correctly.
📺 Source: AI Engineer · Published April 25, 2026
🏷️ Format: Showcase
