# I Turned Coding Agents Into a Strategy Game — Ido Salomon, AgentCraft

## Executive summary

Ido Salomon introduces AgentCraft, a game-inspired orchestrator designed to manage the complexity of large-scale AI agent deployments. The core premise is that while agents are highly capable, the human effort required to steer, direct, and review them at scale creates a bottleneck. AgentCraft addresses this by providing enhanced visibility (mapping the file system and work progress), increasing autonomy (via orchestrators and loops), and enabling seamless collaboration in shared, multi-agent workspaces.

## Key takeaways

- The Human Bottleneck in Agent Deployment: Managing multiple agents is not simple; it requires constant human steering, direction, and review, which becomes exhausting at scale. The skills needed to manage this complexity are analogous to those used in strategy video games like Warcraft or StarCraft.
- AgentCraft: A Game-Inspired Orchestrator: AgentCraft visualizes agents as units on a map where the file system acts as the terrain. It provides integrated functionalities like plugins, skills, terminal, and Git, allowing all work to be managed within a single space.
- Raising the Ceiling: Visibility and Autonomy: Visibility is achieved by projecting the file system onto the map and using heatmaps to visualize where agent work is concentrated. Autonomy is enhanced through 'orchestrators' and 'loops,' which allow agents to execute complex tasks in isolation without constant human intervention.
- Lowering the Floor: Accessibility: To make agents accessible to non-power users (the 90%), a simpler, mobile-game-style mode (potentially 'Loopers') is being developed, focusing on basic, high-level prompts rather than deep file-level granularity.

## Technical details

- Agent Representation: Agents (e.g., cloud code or openclaw agents) are represented as units on a map and can be spawned or detected on the device. They support multimodality, accepting prompts via text and voice.
- Workflow Visualization: The file system can be projected onto the map, allowing users to see exactly which agent is working on which directory or file, visualized as runes. Heatmaps are used to identify areas of high activity.
- Review and Validation: The 'review kit' addresses the difficulty of reviewing multiple changes by providing both file-by-file change logs and visual evidence (videos or photos) of what was modified.
- Collaboration Model: Shared rooms allow human team members (e.g., a product designer and an engineer) and agents to work side-by-side. This shared workspace is not strictly bound to Git, enabling parallel work and direct handoffs.

## Practical implications

- Redefines agent orchestration by moving from linear command execution to a spatial, game-like workflow management system.
- Facilitates complex, multi-stakeholder projects by providing a shared, visible workspace for human and AI collaboration.
- Suggests a shift in focus for build engineers from managing individual agent prompts to designing robust, autonomous agent loops and orchestrators.

## Topics

AI Agents, Agent Orchestration, Workflow Automation, Build Engineering, CI/CD, Human-Computer Interaction, AgentCraft, MCP-UI

Source: https://www.youtube.com/watch?v=YIVkERhy8xo