AgentCraft: Putting the Orc in Orchestration — Ido Salomon

Key Concepts

• Agent Craft: An orchestration platform designed to manage multiple AI agents by applying principles from Real-Time Strategy (RTS) gaming.
• Orchestration: The process of managing, coordinating, and scaling multiple autonomous agents to perform complex tasks.
• Visibility/Observability: The ability to monitor agent activities, file system interactions, and mission status in real-time.
• Human-Agent Collaboration: A framework where humans and AI agents work in a shared workspace, allowing for hand-offs and synchronized task management.
• Task Decomposition: Breaking down high-level goals into smaller, manageable tasks that agents can execute autonomously.

1. Main Topics and Key Points

Edo Salomon introduces Agent Craft, a tool designed to solve the "bottleneck" problem in AI agent management. While individual agents are powerful, scaling to dozens or hundreds creates a management burden for the human engineer. Salomon argues that the skills required to manage these agents are analogous to those used in RTS games, where players manage multiple units simultaneously.

• The Bottleneck: Humans cannot effectively manage dozens of agents using standard CLI tools; the cognitive load is too high.
• Visualizing the File System: Agent Craft maps the user's file system into a visual interface (represented as "rooms" and "buildings"), allowing users to see exactly which files agents are modifying.
• Proactive Collision Prevention: By visualizing agent activity on a map, the system can detect and prevent "collisions" (multiple agents editing the same file simultaneously).

2. Step-by-Step Methodologies

Salomon outlines a progression for scaling agent productivity:

1. Basic Orchestration: Spawning agents and providing direct prompts via a unified interface.
2. Visibility & Monitoring: Using a side panel for mission status and a visual map for file-level tracking.
3. Task Delegation (Quests): Moving from manual prompting to "Quests," where agents identify and execute their own sub-tasks (e.g., refactoring, testing).
4. Campaign Orchestration: Using containers to allow agents to decompose high-level goals into plans, which the human then reviews rather than babysits.
5. Collaborative Workspaces: Enabling multi-user, multi-agent environments where humans and agents share context and state.

3. Key Arguments and Perspectives

• Gaming as a Blueprint: Salomon posits that the "new" skills required for AI management are actually well-developed in the gaming community. Concepts like "muscle memory" for cycling through units and high-level map awareness are directly transferable to software engineering.
• Shift from Planning to Reviewing: As agents become more autonomous, the human role shifts from writing code to "reviewing bundles" of changes. This allows for higher throughput, as a human can generate ten potential solutions and pick the best one.
• Soft Collaboration: The system facilitates communication not just between humans, but between humans and agents, and agents and agents, ensuring that all parties are aware of who is working on which file.

4. Notable Quotes

• "The role of the engineer to actually go and manage dozens of reckless employees is not typically what we do in most companies. So, we need to somehow find these new potentially new skills to manage all of these agents."
• "We're not actually coding, we're just telling other people to code for us or other agents."

5. Technical Features and Real-World Applications

• Visual Lineage: The system tracks which agent performed which action and when, providing a full audit trail of changes.
• Review Bundles: A feature that allows users to view screenshots and videos of agent-generated changes, enabling rapid assessment without deep-diving into code.
• Cross-Functional Collaboration: Product designers and engineers can share a workspace where they can see each other's agents, allowing for seamless hand-offs between different roles.
• Cron-Job Integration: Agents can be set to run on schedules (e.g., scanning Twitter for ideas and implementing them automatically), moving toward fully autonomous workflows.

Synthesis and Conclusion

The core takeaway is that the future of software development lies in orchestration rather than manual coding. By treating AI agents as units in an RTS game, developers can overcome the cognitive limitations of managing multiple autonomous systems. Agent Craft provides the necessary visibility, collaboration, and autonomy to shift the human role from a "coder" to a "manager of agents," ultimately raising the ceiling of what a single developer can achieve.