The main limit on using many agents is the human effort needed to steer, direct, and review them.
2
AgentCraft uses a strategy-game interface to give people visibility, autonomy, review tools, and shared workspaces for agents.
3
A simpler mobile-game-style mode is intended to make agent orchestration accessible to people who are not power users.
Summary
Ido Salomon argues that people, rather than agents, are currently the bottleneck. Running many agents requires constant steering, direction, and review, which becomes exhausting. He draws on games such as Warcraft, The Sims, and Civilization, where players supervise many units and respond to whatever needs attention. AgentCraft applies those ideas to software work. Agents appear on a map, buildings expose tools such as plugins, skills, a terminal, and Git, and a file system can be projected onto the map. Heat maps, attention jumps, quests, orchestrators, loops, and visual review evidence help people manage more work. Shared rooms let people and agents collaborate across roles. Salomon then shows an early, simpler mode called potentially Loopers, aimed at people who may find the full interface too complex. He is direct that this version is experimental and asks viewers to share what they need.
People are the bottleneck because agent supervision becomes exhausting
Salomon starts with a simple challenge: if agents can generate everything from cat memes to B2B SaaS apps, why does everyone not have an army of them? Spinning up many cloud coding agents is easy to imagine, but each agent still needs someone to steer it, direct it, and review its work. The effort becomes exhausting at scale. He argues that the missing capability is not another agent feature. People need ways to supervise many agents without burning out. His comparison to games sets up the rest of the talk: players already know how to watch multiple units, understand what they are doing, and intervene when needed.
Strategy games provide a model for supervising many agents
Salomon compares agent supervision with games such as Warcraft and The Sims. In those games, people watch multiple units moving around and keep track of what each one is doing. He asks how different that is from inspecting agents. AgentCraft takes this comparison directly into software work. It is a game-inspired orchestrator designed to bring familiar game patterns into collaboration between humans and agents. The goal is to raise the ceiling of what one person can coordinate. Instead of treating each agent as an isolated chat session, the interface gives the work a shared space where agents, tasks, tools, and activity can be observed together.
AgentCraft puts agents and development tools onto a shared map
In AgentCraft, an agent appears as a unit on the map. The system can detect agents on the device or let the user spawn new ones from a craft. Salomon says the agents themselves remain ordinary agents, including agents such as Claude Code or OpenClaw, and users can prompt them through a side panel with voice support. Other map elements appear as buildings. These buildings expose functions such as plugin and skill management, an integrated terminal, and integrated Git. The point is to let people do their work inside one space while keeping the agents visible as active units.
A projected file system and attention controls make agent activity easier to inspect
Visibility is the first part of raising the ceiling. A side panel shows each agent, its task, its last action, and what it is doing now, so the user can see who needs attention. Salomon then projects a file system onto the map. Directories and files become visible as runes, allowing him to see which part of a project an agent is working on and what it changed. The orchestrator can use this information to create heat maps. When something needs attention, pressing the space bar jumps to the nearest item, similar to an RTS game or Civilization. The user can then answer questions or approve plans without searching manually.
Quests reduce the amount of project planning a person must hold in mind
Visibility alone still leaves too much work for the person. Salomon says it is hard to hold a large coding project in mind while deciding what to do next, even when all the agent activity is visible. AgentCraft can ask agents to go through a codebase, identify the next things that need doing, and create tasks. The user can then accept one of these tasks as a quest with a button, after which an agent works on it. This shifts some planning into the system. The person still chooses what to accept, but does not have to formulate every next task while also watching all the other work.
Orchestrators, loops, and isolated execution let agents work with less babysitting
The next problem is that accepting many tasks can create a new supervision burden. Salomon describes software factories, where a user gives an agent a general sense of the desired result. The agent spawns an orchestrator, breaks the work down, and runs it in an isolated local container. The user does not need to babysit each step. Loops provide another form of autonomy. Salomon gives examples such as scanning Twitter for ideas to build or scanning GitHub, then creating features in the background. These patterns let agents continue work from broad instructions, although they create a separate need for reviewing several results at once.
The review kit uses changed files and visual evidence to speed up decisions
Autonomous work creates a review problem when several implementations are ready. AgentCraft's review kit presents the complete set of changes and the files that changed. A reviewer can inspect the work file by file or use visual evidence, including videos or photos of what changed. Salomon also describes running several instances in parallel and choosing the best implementation. This gives the user different levels of review, from detailed file inspection to a quicker view of the resulting behavior. In his progression, visibility and autonomy make larger-scale work possible, but review remains the point where the person has to make decisions.
Shared rooms let people and agents work together across roles
Salomon does not want one person to be the only reviewer or the only person talking to agents. AgentCraft can host shared rooms locally, with tunnels allowing other people to join. In his example, a product designer works in the room while the engineer follows the design and implements it. The work can be handed off or forked in the shared workspace, rather than being limited to a Git-centered exchange. People and agents also have a shared chat or notice board, so participants can see what others are doing. Salomon says the room can include different devices and Telegram, giving collaborators more ways to stay involved.
A simpler mode could lower the entry barrier for people who find the full interface too complex
After describing how AgentCraft raises the ceiling for power users, Salomon turns to lowering the floor. He says some surprising users, including children and people with no prior interest in agents or productivity tools, enjoyed orchestrating agents through the game-like interface. He wants to reach the much larger group that may find a Warcraft-style interface too complicated. The early concept, currently called potentially Loopers, uses a more mobile-game-like experience. Projects could remain available for repeated work, while loops and more autonomous agents handle the details. A person might prompt the system, receive a result, and avoid seeing every file unless they choose to inspect it. Salomon says the design is still experimental and should be customizable because the same interface will not suit everyone.
"I took this concept too far and I created AgentCraft, which is a game-inspired orchestrator that tries to bring all that good stuff together into work."02:14
Who should watch
You are running several coding agents and spend too much time switching between prompts, status checks, and reviews.
You want to understand how a game-like interface could organize agent activity, tasks, autonomous loops, and human approval.
You are designing agent tools for people who are not already power users and want ideas for making the first experience less intimidating.
