AgentCraft

Salomon opens with an imagined setup of 25 agents, then argues that steering and reviewing them makes human attention the bottleneck. In his demonstration, agents appear as units; buildings expose skills, plugins, a terminal and Git. A side panel shows tasks, recent actions and current activity. Projecting the file system onto the map connects agents to directories and files, with activity histories and heat maps showing where work occurs. Pressing the space bar moves to an item needing attention, such as a question or plan approval. These controls connect observation to intervention, but Salomon explicitly notes that watching file and tool activity remains exhausting.

The recording describes agents inspecting a codebase and proposing tasks as quests that a person accepts. A general goal can instead launch an orchestrator that decomposes and runs the work; recurring loops can scan Twitter or GitHub for work to undertake. Salomon describes the demonstrated execution environment as a local container. Separately, the official site advertises Docker or Apple Containers, full network isolation and separate browser sessions for agents running development servers on the same port. Those are site claims, not isolation properties established by the recording. Delegation also concentrates review work: Salomon says even five simultaneous results can be difficult to assess. His review kit offers changed files alongside images or videos, and supports comparing parallel implementations.

Salomon argues that the person directing agents may not be the appropriate reviewer for every task. He demonstrates locally hosted rooms that others can join through tunnels. In his example, a designer’s work becomes visible on an engineer’s map, allowing implementation to begin while design continues. Handoffs, forks, chat and a notice board communicate what people and agents are doing; Git remains available. The official site calls its multiplayer feature Alliance Hall and describes remote agents from other machines appearing on the shared map. It also lists a mobile PWA, push notifications with quick replies, and Telegram/Discord channels. The recording discusses mobile and Telegram access, without establishing every capability listed on the site.

The later demonstration introduces a very early concept tentatively called Loopers. It replaces required file-level supervision with recurring projects, autonomous loops and a simpler prompt-to-result interaction, while leaving process inspection available. The tradeoff is greater reliance on autonomous execution in exchange for less detail to manage. Salomon’s stories about children and former strategy-game players using agents are reported feedback, not measured adoption findings; his roughly 90% estimate of people needing a lower entry point is likewise an approximation. He acknowledges that a game interface can feel accessible to one person and difficult to another, making customization part of the experiment.

The talk ends by distinguishing available AgentCraft from the experimental simpler concept. The official site supplies the launch command npx @idosal/agentcraft; Salomon seeks expressions of interest and feedback for the tentative Loopers direction. His closing argument is that familiar skills from games and other activities can help people supervise agents, and he invites listeners to apply what fits their own work. This is a proposed approach to reducing the human bottleneck, not a demonstrated elimination of oversight or proof that one game metaphor works for everyone.